Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
| 2 | * INET An implementation of the TCP/IP protocol suite for the LINUX |
| 3 | * operating system. INET is implemented using the BSD Socket |
| 4 | * interface as the means of communication with the user level. |
| 5 | * |
| 6 | * Implementation of the Transmission Control Protocol(TCP). |
| 7 | * |
Jesper Juhl | 02c30a8 | 2005-05-05 16:16:16 -0700 | [diff] [blame] | 8 | * Authors: Ross Biro |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 9 | * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> |
| 10 | * Mark Evans, <evansmp@uhura.aston.ac.uk> |
| 11 | * Corey Minyard <wf-rch!minyard@relay.EU.net> |
| 12 | * Florian La Roche, <flla@stud.uni-sb.de> |
| 13 | * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> |
| 14 | * Linus Torvalds, <torvalds@cs.helsinki.fi> |
| 15 | * Alan Cox, <gw4pts@gw4pts.ampr.org> |
| 16 | * Matthew Dillon, <dillon@apollo.west.oic.com> |
| 17 | * Arnt Gulbrandsen, <agulbra@nvg.unit.no> |
| 18 | * Jorge Cwik, <jorge@laser.satlink.net> |
| 19 | */ |
| 20 | |
| 21 | /* |
| 22 | * Changes: |
| 23 | * Pedro Roque : Fast Retransmit/Recovery. |
| 24 | * Two receive queues. |
| 25 | * Retransmit queue handled by TCP. |
| 26 | * Better retransmit timer handling. |
| 27 | * New congestion avoidance. |
| 28 | * Header prediction. |
| 29 | * Variable renaming. |
| 30 | * |
| 31 | * Eric : Fast Retransmit. |
| 32 | * Randy Scott : MSS option defines. |
| 33 | * Eric Schenk : Fixes to slow start algorithm. |
| 34 | * Eric Schenk : Yet another double ACK bug. |
| 35 | * Eric Schenk : Delayed ACK bug fixes. |
| 36 | * Eric Schenk : Floyd style fast retrans war avoidance. |
| 37 | * David S. Miller : Don't allow zero congestion window. |
| 38 | * Eric Schenk : Fix retransmitter so that it sends |
| 39 | * next packet on ack of previous packet. |
| 40 | * Andi Kleen : Moved open_request checking here |
| 41 | * and process RSTs for open_requests. |
| 42 | * Andi Kleen : Better prune_queue, and other fixes. |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 43 | * Andrey Savochkin: Fix RTT measurements in the presence of |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 44 | * timestamps. |
| 45 | * Andrey Savochkin: Check sequence numbers correctly when |
| 46 | * removing SACKs due to in sequence incoming |
| 47 | * data segments. |
| 48 | * Andi Kleen: Make sure we never ack data there is not |
| 49 | * enough room for. Also make this condition |
| 50 | * a fatal error if it might still happen. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 51 | * Andi Kleen: Add tcp_measure_rcv_mss to make |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 52 | * connections with MSS<min(MTU,ann. MSS) |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 53 | * work without delayed acks. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 54 | * Andi Kleen: Process packets with PSH set in the |
| 55 | * fast path. |
| 56 | * J Hadi Salim: ECN support |
| 57 | * Andrei Gurtov, |
| 58 | * Pasi Sarolahti, |
| 59 | * Panu Kuhlberg: Experimental audit of TCP (re)transmission |
| 60 | * engine. Lots of bugs are found. |
| 61 | * Pasi Sarolahti: F-RTO for dealing with spurious RTOs |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 62 | */ |
| 63 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 64 | #include <linux/mm.h> |
| 65 | #include <linux/module.h> |
| 66 | #include <linux/sysctl.h> |
Ilpo Järvinen | a0bffff | 2009-03-21 13:36:17 -0700 | [diff] [blame] | 67 | #include <linux/kernel.h> |
Satoru SATOH | 5ffc02a | 2008-05-04 22:14:42 -0700 | [diff] [blame] | 68 | #include <net/dst.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 69 | #include <net/tcp.h> |
| 70 | #include <net/inet_common.h> |
| 71 | #include <linux/ipsec.h> |
| 72 | #include <asm/unaligned.h> |
Chris Leech | 1a2449a | 2006-05-23 18:05:53 -0700 | [diff] [blame] | 73 | #include <net/netdma.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 74 | |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 75 | int sysctl_tcp_timestamps __read_mostly = 1; |
| 76 | int sysctl_tcp_window_scaling __read_mostly = 1; |
| 77 | int sysctl_tcp_sack __read_mostly = 1; |
| 78 | int sysctl_tcp_fack __read_mostly = 1; |
| 79 | int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH; |
Ilpo Järvinen | 255cac9 | 2009-05-04 11:07:36 -0700 | [diff] [blame] | 80 | int sysctl_tcp_ecn __read_mostly = 2; |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 81 | int sysctl_tcp_dsack __read_mostly = 1; |
| 82 | int sysctl_tcp_app_win __read_mostly = 31; |
| 83 | int sysctl_tcp_adv_win_scale __read_mostly = 2; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 84 | |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 85 | int sysctl_tcp_stdurg __read_mostly; |
| 86 | int sysctl_tcp_rfc1337 __read_mostly; |
| 87 | int sysctl_tcp_max_orphans __read_mostly = NR_FILE; |
Ilpo Järvinen | c96fd3d | 2007-09-20 11:36:37 -0700 | [diff] [blame] | 88 | int sysctl_tcp_frto __read_mostly = 2; |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 89 | int sysctl_tcp_frto_response __read_mostly; |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 90 | int sysctl_tcp_nometrics_save __read_mostly; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 91 | |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 92 | int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; |
| 93 | int sysctl_tcp_abc __read_mostly; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 94 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 95 | #define FLAG_DATA 0x01 /* Incoming frame contained data. */ |
| 96 | #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ |
| 97 | #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ |
| 98 | #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ |
| 99 | #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ |
| 100 | #define FLAG_DATA_SACKED 0x20 /* New SACK. */ |
| 101 | #define FLAG_ECE 0x40 /* ECE in this ACK */ |
| 102 | #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */ |
| 103 | #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ |
Ilpo Järvinen | 4dc2665 | 2007-02-21 23:16:11 -0800 | [diff] [blame] | 104 | #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */ |
Ilpo Järvinen | 2e60529 | 2007-08-02 19:46:58 -0700 | [diff] [blame] | 105 | #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ |
Ryousei Takano | 564262c1 | 2007-10-25 23:03:52 -0700 | [diff] [blame] | 106 | #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ |
Ilpo Järvinen | 009a2e3 | 2007-09-20 11:34:38 -0700 | [diff] [blame] | 107 | #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */ |
Ilpo Järvinen | cadbd03 | 2007-12-31 04:49:21 -0800 | [diff] [blame] | 108 | #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 109 | |
| 110 | #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) |
| 111 | #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) |
| 112 | #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) |
| 113 | #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) |
Ilpo Järvinen | 2e60529 | 2007-08-02 19:46:58 -0700 | [diff] [blame] | 114 | #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 115 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 116 | #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 117 | #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 118 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 119 | /* Adapt the MSS value used to make delayed ack decision to the |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 120 | * real world. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 121 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 122 | static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 123 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 124 | struct inet_connection_sock *icsk = inet_csk(sk); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 125 | const unsigned int lss = icsk->icsk_ack.last_seg_size; |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 126 | unsigned int len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 127 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 128 | icsk->icsk_ack.last_seg_size = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 129 | |
| 130 | /* skb->len may jitter because of SACKs, even if peer |
| 131 | * sends good full-sized frames. |
| 132 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 133 | len = skb_shinfo(skb)->gso_size ? : skb->len; |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 134 | if (len >= icsk->icsk_ack.rcv_mss) { |
| 135 | icsk->icsk_ack.rcv_mss = len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 136 | } else { |
| 137 | /* Otherwise, we make more careful check taking into account, |
| 138 | * that SACKs block is variable. |
| 139 | * |
| 140 | * "len" is invariant segment length, including TCP header. |
| 141 | */ |
Arnaldo Carvalho de Melo | 9c70220 | 2007-04-25 18:04:18 -0700 | [diff] [blame] | 142 | len += skb->data - skb_transport_header(skb); |
William Allen Simpson | bee7ca9 | 2009-11-10 09:51:18 +0000 | [diff] [blame] | 143 | if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 144 | /* If PSH is not set, packet should be |
| 145 | * full sized, provided peer TCP is not badly broken. |
| 146 | * This observation (if it is correct 8)) allows |
| 147 | * to handle super-low mtu links fairly. |
| 148 | */ |
| 149 | (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 150 | !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 151 | /* Subtract also invariant (if peer is RFC compliant), |
| 152 | * tcp header plus fixed timestamp option length. |
| 153 | * Resulting "len" is MSS free of SACK jitter. |
| 154 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 155 | len -= tcp_sk(sk)->tcp_header_len; |
| 156 | icsk->icsk_ack.last_seg_size = len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 157 | if (len == lss) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 158 | icsk->icsk_ack.rcv_mss = len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 159 | return; |
| 160 | } |
| 161 | } |
Alexey Kuznetsov | 1ef9696 | 2006-09-19 12:52:50 -0700 | [diff] [blame] | 162 | if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) |
| 163 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 164 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 165 | } |
| 166 | } |
| 167 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 168 | static void tcp_incr_quickack(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 169 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 170 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 171 | unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 172 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 173 | if (quickacks == 0) |
| 174 | quickacks = 2; |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 175 | if (quickacks > icsk->icsk_ack.quick) |
| 176 | icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 177 | } |
| 178 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 179 | void tcp_enter_quickack_mode(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 180 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 181 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 182 | tcp_incr_quickack(sk); |
| 183 | icsk->icsk_ack.pingpong = 0; |
| 184 | icsk->icsk_ack.ato = TCP_ATO_MIN; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 185 | } |
| 186 | |
| 187 | /* Send ACKs quickly, if "quick" count is not exhausted |
| 188 | * and the session is not interactive. |
| 189 | */ |
| 190 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 191 | static inline int tcp_in_quickack_mode(const struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 192 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 193 | const struct inet_connection_sock *icsk = inet_csk(sk); |
| 194 | return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 195 | } |
| 196 | |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 197 | static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp) |
| 198 | { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 199 | if (tp->ecn_flags & TCP_ECN_OK) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 200 | tp->ecn_flags |= TCP_ECN_QUEUE_CWR; |
| 201 | } |
| 202 | |
| 203 | static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb) |
| 204 | { |
| 205 | if (tcp_hdr(skb)->cwr) |
| 206 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; |
| 207 | } |
| 208 | |
| 209 | static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp) |
| 210 | { |
| 211 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; |
| 212 | } |
| 213 | |
| 214 | static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb) |
| 215 | { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 216 | if (tp->ecn_flags & TCP_ECN_OK) { |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 217 | if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags)) |
| 218 | tp->ecn_flags |= TCP_ECN_DEMAND_CWR; |
| 219 | /* Funny extension: if ECT is not set on a segment, |
| 220 | * it is surely retransmit. It is not in ECN RFC, |
| 221 | * but Linux follows this rule. */ |
| 222 | else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags))) |
| 223 | tcp_enter_quickack_mode((struct sock *)tp); |
| 224 | } |
| 225 | } |
| 226 | |
| 227 | static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th) |
| 228 | { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 229 | if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 230 | tp->ecn_flags &= ~TCP_ECN_OK; |
| 231 | } |
| 232 | |
| 233 | static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th) |
| 234 | { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 235 | if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 236 | tp->ecn_flags &= ~TCP_ECN_OK; |
| 237 | } |
| 238 | |
| 239 | static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th) |
| 240 | { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 241 | if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 242 | return 1; |
| 243 | return 0; |
| 244 | } |
| 245 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 246 | /* Buffer size and advertised window tuning. |
| 247 | * |
| 248 | * 1. Tuning sk->sk_sndbuf, when connection enters established state. |
| 249 | */ |
| 250 | |
| 251 | static void tcp_fixup_sndbuf(struct sock *sk) |
| 252 | { |
| 253 | int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 + |
| 254 | sizeof(struct sk_buff); |
| 255 | |
| 256 | if (sk->sk_sndbuf < 3 * sndmem) |
| 257 | sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]); |
| 258 | } |
| 259 | |
| 260 | /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) |
| 261 | * |
| 262 | * All tcp_full_space() is split to two parts: "network" buffer, allocated |
| 263 | * forward and advertised in receiver window (tp->rcv_wnd) and |
| 264 | * "application buffer", required to isolate scheduling/application |
| 265 | * latencies from network. |
| 266 | * window_clamp is maximal advertised window. It can be less than |
| 267 | * tcp_full_space(), in this case tcp_full_space() - window_clamp |
| 268 | * is reserved for "application" buffer. The less window_clamp is |
| 269 | * the smoother our behaviour from viewpoint of network, but the lower |
| 270 | * throughput and the higher sensitivity of the connection to losses. 8) |
| 271 | * |
| 272 | * rcv_ssthresh is more strict window_clamp used at "slow start" |
| 273 | * phase to predict further behaviour of this connection. |
| 274 | * It is used for two goals: |
| 275 | * - to enforce header prediction at sender, even when application |
| 276 | * requires some significant "application buffer". It is check #1. |
| 277 | * - to prevent pruning of receive queue because of misprediction |
| 278 | * of receiver window. Check #2. |
| 279 | * |
| 280 | * The scheme does not work when sender sends good segments opening |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 281 | * window and then starts to feed us spaghetti. But it should work |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 282 | * in common situations. Otherwise, we have to rely on queue collapsing. |
| 283 | */ |
| 284 | |
| 285 | /* Slow part of check#2. */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 286 | static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 287 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 288 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 289 | /* Optimize this! */ |
Eric Dumazet | dfd4f0a | 2007-12-21 06:07:53 -0800 | [diff] [blame] | 290 | int truesize = tcp_win_from_space(skb->truesize) >> 1; |
| 291 | int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 292 | |
| 293 | while (tp->rcv_ssthresh <= window) { |
| 294 | if (truesize <= skb->len) |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 295 | return 2 * inet_csk(sk)->icsk_ack.rcv_mss; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 296 | |
| 297 | truesize >>= 1; |
| 298 | window >>= 1; |
| 299 | } |
| 300 | return 0; |
| 301 | } |
| 302 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 303 | static void tcp_grow_window(struct sock *sk, struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 304 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 305 | struct tcp_sock *tp = tcp_sk(sk); |
| 306 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 307 | /* Check #1 */ |
| 308 | if (tp->rcv_ssthresh < tp->window_clamp && |
| 309 | (int)tp->rcv_ssthresh < tcp_space(sk) && |
| 310 | !tcp_memory_pressure) { |
| 311 | int incr; |
| 312 | |
| 313 | /* Check #2. Increase window, if skb with such overhead |
| 314 | * will fit to rcvbuf in future. |
| 315 | */ |
| 316 | if (tcp_win_from_space(skb->truesize) <= skb->len) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 317 | incr = 2 * tp->advmss; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 318 | else |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 319 | incr = __tcp_grow_window(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 320 | |
| 321 | if (incr) { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 322 | tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, |
| 323 | tp->window_clamp); |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 324 | inet_csk(sk)->icsk_ack.quick |= 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 325 | } |
| 326 | } |
| 327 | } |
| 328 | |
| 329 | /* 3. Tuning rcvbuf, when connection enters established state. */ |
| 330 | |
| 331 | static void tcp_fixup_rcvbuf(struct sock *sk) |
| 332 | { |
| 333 | struct tcp_sock *tp = tcp_sk(sk); |
| 334 | int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); |
| 335 | |
| 336 | /* Try to select rcvbuf so that 4 mss-sized segments |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 337 | * will fit to window and corresponding skbs will fit to our rcvbuf. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 338 | * (was 3; 4 is minimum to allow fast retransmit to work.) |
| 339 | */ |
| 340 | while (tcp_win_from_space(rcvmem) < tp->advmss) |
| 341 | rcvmem += 128; |
| 342 | if (sk->sk_rcvbuf < 4 * rcvmem) |
| 343 | sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]); |
| 344 | } |
| 345 | |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 346 | /* 4. Try to fixup all. It is made immediately after connection enters |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 347 | * established state. |
| 348 | */ |
| 349 | static void tcp_init_buffer_space(struct sock *sk) |
| 350 | { |
| 351 | struct tcp_sock *tp = tcp_sk(sk); |
| 352 | int maxwin; |
| 353 | |
| 354 | if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) |
| 355 | tcp_fixup_rcvbuf(sk); |
| 356 | if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) |
| 357 | tcp_fixup_sndbuf(sk); |
| 358 | |
| 359 | tp->rcvq_space.space = tp->rcv_wnd; |
| 360 | |
| 361 | maxwin = tcp_full_space(sk); |
| 362 | |
| 363 | if (tp->window_clamp >= maxwin) { |
| 364 | tp->window_clamp = maxwin; |
| 365 | |
| 366 | if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) |
| 367 | tp->window_clamp = max(maxwin - |
| 368 | (maxwin >> sysctl_tcp_app_win), |
| 369 | 4 * tp->advmss); |
| 370 | } |
| 371 | |
| 372 | /* Force reservation of one segment. */ |
| 373 | if (sysctl_tcp_app_win && |
| 374 | tp->window_clamp > 2 * tp->advmss && |
| 375 | tp->window_clamp + tp->advmss > maxwin) |
| 376 | tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); |
| 377 | |
| 378 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); |
| 379 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 380 | } |
| 381 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 382 | /* 5. Recalculate window clamp after socket hit its memory bounds. */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 383 | static void tcp_clamp_window(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 384 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 385 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 386 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 387 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 388 | icsk->icsk_ack.quick = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 389 | |
John Heffner | 326f36e | 2005-11-10 17:11:48 -0800 | [diff] [blame] | 390 | if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && |
| 391 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && |
| 392 | !tcp_memory_pressure && |
| 393 | atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) { |
| 394 | sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), |
| 395 | sysctl_tcp_rmem[2]); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 396 | } |
John Heffner | 326f36e | 2005-11-10 17:11:48 -0800 | [diff] [blame] | 397 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 398 | tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 399 | } |
| 400 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 401 | /* Initialize RCV_MSS value. |
| 402 | * RCV_MSS is an our guess about MSS used by the peer. |
| 403 | * We haven't any direct information about the MSS. |
| 404 | * It's better to underestimate the RCV_MSS rather than overestimate. |
| 405 | * Overestimations make us ACKing less frequently than needed. |
| 406 | * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). |
| 407 | */ |
| 408 | void tcp_initialize_rcv_mss(struct sock *sk) |
| 409 | { |
| 410 | struct tcp_sock *tp = tcp_sk(sk); |
| 411 | unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); |
| 412 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 413 | hint = min(hint, tp->rcv_wnd / 2); |
William Allen Simpson | bee7ca9 | 2009-11-10 09:51:18 +0000 | [diff] [blame] | 414 | hint = min(hint, TCP_MSS_DEFAULT); |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 415 | hint = max(hint, TCP_MIN_MSS); |
| 416 | |
| 417 | inet_csk(sk)->icsk_ack.rcv_mss = hint; |
| 418 | } |
| 419 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 420 | /* Receiver "autotuning" code. |
| 421 | * |
| 422 | * The algorithm for RTT estimation w/o timestamps is based on |
| 423 | * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. |
| 424 | * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps> |
| 425 | * |
| 426 | * More detail on this code can be found at |
| 427 | * <http://www.psc.edu/~jheffner/senior_thesis.ps>, |
| 428 | * though this reference is out of date. A new paper |
| 429 | * is pending. |
| 430 | */ |
| 431 | static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) |
| 432 | { |
| 433 | u32 new_sample = tp->rcv_rtt_est.rtt; |
| 434 | long m = sample; |
| 435 | |
| 436 | if (m == 0) |
| 437 | m = 1; |
| 438 | |
| 439 | if (new_sample != 0) { |
| 440 | /* If we sample in larger samples in the non-timestamp |
| 441 | * case, we could grossly overestimate the RTT especially |
| 442 | * with chatty applications or bulk transfer apps which |
| 443 | * are stalled on filesystem I/O. |
| 444 | * |
| 445 | * Also, since we are only going for a minimum in the |
Stephen Hemminger | 31f3426 | 2005-11-15 15:17:10 -0800 | [diff] [blame] | 446 | * non-timestamp case, we do not smooth things out |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 447 | * else with timestamps disabled convergence takes too |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 448 | * long. |
| 449 | */ |
| 450 | if (!win_dep) { |
| 451 | m -= (new_sample >> 3); |
| 452 | new_sample += m; |
| 453 | } else if (m < new_sample) |
| 454 | new_sample = m << 3; |
| 455 | } else { |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 456 | /* No previous measure. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 457 | new_sample = m << 3; |
| 458 | } |
| 459 | |
| 460 | if (tp->rcv_rtt_est.rtt != new_sample) |
| 461 | tp->rcv_rtt_est.rtt = new_sample; |
| 462 | } |
| 463 | |
| 464 | static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) |
| 465 | { |
| 466 | if (tp->rcv_rtt_est.time == 0) |
| 467 | goto new_measure; |
| 468 | if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) |
| 469 | return; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 470 | tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 471 | |
| 472 | new_measure: |
| 473 | tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; |
| 474 | tp->rcv_rtt_est.time = tcp_time_stamp; |
| 475 | } |
| 476 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 477 | static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, |
| 478 | const struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 479 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 480 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 481 | if (tp->rx_opt.rcv_tsecr && |
| 482 | (TCP_SKB_CB(skb)->end_seq - |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 483 | TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 484 | tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); |
| 485 | } |
| 486 | |
| 487 | /* |
| 488 | * This function should be called every time data is copied to user space. |
| 489 | * It calculates the appropriate TCP receive buffer space. |
| 490 | */ |
| 491 | void tcp_rcv_space_adjust(struct sock *sk) |
| 492 | { |
| 493 | struct tcp_sock *tp = tcp_sk(sk); |
| 494 | int time; |
| 495 | int space; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 496 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 497 | if (tp->rcvq_space.time == 0) |
| 498 | goto new_measure; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 499 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 500 | time = tcp_time_stamp - tp->rcvq_space.time; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 501 | if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 502 | return; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 503 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 504 | space = 2 * (tp->copied_seq - tp->rcvq_space.seq); |
| 505 | |
| 506 | space = max(tp->rcvq_space.space, space); |
| 507 | |
| 508 | if (tp->rcvq_space.space != space) { |
| 509 | int rcvmem; |
| 510 | |
| 511 | tp->rcvq_space.space = space; |
| 512 | |
John Heffner | 6fcf941 | 2006-02-09 17:06:57 -0800 | [diff] [blame] | 513 | if (sysctl_tcp_moderate_rcvbuf && |
| 514 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 515 | int new_clamp = space; |
| 516 | |
| 517 | /* Receive space grows, normalize in order to |
| 518 | * take into account packet headers and sk_buff |
| 519 | * structure overhead. |
| 520 | */ |
| 521 | space /= tp->advmss; |
| 522 | if (!space) |
| 523 | space = 1; |
| 524 | rcvmem = (tp->advmss + MAX_TCP_HEADER + |
| 525 | 16 + sizeof(struct sk_buff)); |
| 526 | while (tcp_win_from_space(rcvmem) < tp->advmss) |
| 527 | rcvmem += 128; |
| 528 | space *= rcvmem; |
| 529 | space = min(space, sysctl_tcp_rmem[2]); |
| 530 | if (space > sk->sk_rcvbuf) { |
| 531 | sk->sk_rcvbuf = space; |
| 532 | |
| 533 | /* Make the window clamp follow along. */ |
| 534 | tp->window_clamp = new_clamp; |
| 535 | } |
| 536 | } |
| 537 | } |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 538 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 539 | new_measure: |
| 540 | tp->rcvq_space.seq = tp->copied_seq; |
| 541 | tp->rcvq_space.time = tcp_time_stamp; |
| 542 | } |
| 543 | |
| 544 | /* There is something which you must keep in mind when you analyze the |
| 545 | * behavior of the tp->ato delayed ack timeout interval. When a |
| 546 | * connection starts up, we want to ack as quickly as possible. The |
| 547 | * problem is that "good" TCP's do slow start at the beginning of data |
| 548 | * transmission. The means that until we send the first few ACK's the |
| 549 | * sender will sit on his end and only queue most of his data, because |
| 550 | * he can only send snd_cwnd unacked packets at any given time. For |
| 551 | * each ACK we send, he increments snd_cwnd and transmits more of his |
| 552 | * queue. -DaveM |
| 553 | */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 554 | static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 555 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 556 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 557 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 558 | u32 now; |
| 559 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 560 | inet_csk_schedule_ack(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 561 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 562 | tcp_measure_rcv_mss(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 563 | |
| 564 | tcp_rcv_rtt_measure(tp); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 565 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 566 | now = tcp_time_stamp; |
| 567 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 568 | if (!icsk->icsk_ack.ato) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 569 | /* The _first_ data packet received, initialize |
| 570 | * delayed ACK engine. |
| 571 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 572 | tcp_incr_quickack(sk); |
| 573 | icsk->icsk_ack.ato = TCP_ATO_MIN; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 574 | } else { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 575 | int m = now - icsk->icsk_ack.lrcvtime; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 576 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 577 | if (m <= TCP_ATO_MIN / 2) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 578 | /* The fastest case is the first. */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 579 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; |
| 580 | } else if (m < icsk->icsk_ack.ato) { |
| 581 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; |
| 582 | if (icsk->icsk_ack.ato > icsk->icsk_rto) |
| 583 | icsk->icsk_ack.ato = icsk->icsk_rto; |
| 584 | } else if (m > icsk->icsk_rto) { |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 585 | /* Too long gap. Apparently sender failed to |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 586 | * restart window, so that we send ACKs quickly. |
| 587 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 588 | tcp_incr_quickack(sk); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 589 | sk_mem_reclaim(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 590 | } |
| 591 | } |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 592 | icsk->icsk_ack.lrcvtime = now; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 593 | |
| 594 | TCP_ECN_check_ce(tp, skb); |
| 595 | |
| 596 | if (skb->len >= 128) |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 597 | tcp_grow_window(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 598 | } |
| 599 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 600 | /* Called to compute a smoothed rtt estimate. The data fed to this |
| 601 | * routine either comes from timestamps, or from segments that were |
| 602 | * known _not_ to have been retransmitted [see Karn/Partridge |
| 603 | * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 |
| 604 | * piece by Van Jacobson. |
| 605 | * NOTE: the next three routines used to be one big routine. |
| 606 | * To save cycles in the RFC 1323 implementation it was better to break |
| 607 | * it up into three procedures. -- erics |
| 608 | */ |
Stephen Hemminger | 2d2abba | 2005-11-10 16:56:12 -0800 | [diff] [blame] | 609 | static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 610 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 611 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 612 | long m = mrtt; /* RTT */ |
| 613 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 614 | /* The following amusing code comes from Jacobson's |
| 615 | * article in SIGCOMM '88. Note that rtt and mdev |
| 616 | * are scaled versions of rtt and mean deviation. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 617 | * This is designed to be as fast as possible |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 618 | * m stands for "measurement". |
| 619 | * |
| 620 | * On a 1990 paper the rto value is changed to: |
| 621 | * RTO = rtt + 4 * mdev |
| 622 | * |
| 623 | * Funny. This algorithm seems to be very broken. |
| 624 | * These formulae increase RTO, when it should be decreased, increase |
Stephen Hemminger | 31f3426 | 2005-11-15 15:17:10 -0800 | [diff] [blame] | 625 | * too slowly, when it should be increased quickly, decrease too quickly |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 626 | * etc. I guess in BSD RTO takes ONE value, so that it is absolutely |
| 627 | * does not matter how to _calculate_ it. Seems, it was trap |
| 628 | * that VJ failed to avoid. 8) |
| 629 | */ |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 630 | if (m == 0) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 631 | m = 1; |
| 632 | if (tp->srtt != 0) { |
| 633 | m -= (tp->srtt >> 3); /* m is now error in rtt est */ |
| 634 | tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ |
| 635 | if (m < 0) { |
| 636 | m = -m; /* m is now abs(error) */ |
| 637 | m -= (tp->mdev >> 2); /* similar update on mdev */ |
| 638 | /* This is similar to one of Eifel findings. |
| 639 | * Eifel blocks mdev updates when rtt decreases. |
| 640 | * This solution is a bit different: we use finer gain |
| 641 | * for mdev in this case (alpha*beta). |
| 642 | * Like Eifel it also prevents growth of rto, |
| 643 | * but also it limits too fast rto decreases, |
| 644 | * happening in pure Eifel. |
| 645 | */ |
| 646 | if (m > 0) |
| 647 | m >>= 3; |
| 648 | } else { |
| 649 | m -= (tp->mdev >> 2); /* similar update on mdev */ |
| 650 | } |
| 651 | tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ |
| 652 | if (tp->mdev > tp->mdev_max) { |
| 653 | tp->mdev_max = tp->mdev; |
| 654 | if (tp->mdev_max > tp->rttvar) |
| 655 | tp->rttvar = tp->mdev_max; |
| 656 | } |
| 657 | if (after(tp->snd_una, tp->rtt_seq)) { |
| 658 | if (tp->mdev_max < tp->rttvar) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 659 | tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 660 | tp->rtt_seq = tp->snd_nxt; |
David S. Miller | 05bb1fa | 2007-08-30 22:10:28 -0700 | [diff] [blame] | 661 | tp->mdev_max = tcp_rto_min(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 662 | } |
| 663 | } else { |
| 664 | /* no previous measure. */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 665 | tp->srtt = m << 3; /* take the measured time to be rtt */ |
| 666 | tp->mdev = m << 1; /* make sure rto = 3*rtt */ |
David S. Miller | 05bb1fa | 2007-08-30 22:10:28 -0700 | [diff] [blame] | 667 | tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 668 | tp->rtt_seq = tp->snd_nxt; |
| 669 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 670 | } |
| 671 | |
| 672 | /* Calculate rto without backoff. This is the second half of Van Jacobson's |
| 673 | * routine referred to above. |
| 674 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 675 | static inline void tcp_set_rto(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 676 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 677 | const struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 678 | /* Old crap is replaced with new one. 8) |
| 679 | * |
| 680 | * More seriously: |
| 681 | * 1. If rtt variance happened to be less 50msec, it is hallucination. |
| 682 | * It cannot be less due to utterly erratic ACK generation made |
| 683 | * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ |
| 684 | * to do with delayed acks, because at cwnd>2 true delack timeout |
| 685 | * is invisible. Actually, Linux-2.4 also generates erratic |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 686 | * ACKs in some circumstances. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 687 | */ |
Damian Lukowski | f1ecd5d | 2009-08-26 00:16:31 +0000 | [diff] [blame] | 688 | inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 689 | |
| 690 | /* 2. Fixups made earlier cannot be right. |
| 691 | * If we do not estimate RTO correctly without them, |
| 692 | * all the algo is pure shit and should be replaced |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 693 | * with correct one. It is exactly, which we pretend to do. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 694 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 695 | |
Ilpo Järvinen | ee6aac5 | 2008-12-05 22:43:08 -0800 | [diff] [blame] | 696 | /* NOTE: clamping at TCP_RTO_MIN is not required, current algo |
| 697 | * guarantees that rto is higher. |
| 698 | */ |
Damian Lukowski | f1ecd5d | 2009-08-26 00:16:31 +0000 | [diff] [blame] | 699 | tcp_bound_rto(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 700 | } |
| 701 | |
| 702 | /* Save metrics learned by this TCP session. |
| 703 | This function is called only, when TCP finishes successfully |
| 704 | i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. |
| 705 | */ |
| 706 | void tcp_update_metrics(struct sock *sk) |
| 707 | { |
| 708 | struct tcp_sock *tp = tcp_sk(sk); |
| 709 | struct dst_entry *dst = __sk_dst_get(sk); |
| 710 | |
| 711 | if (sysctl_tcp_nometrics_save) |
| 712 | return; |
| 713 | |
| 714 | dst_confirm(dst); |
| 715 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 716 | if (dst && (dst->flags & DST_HOST)) { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 717 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 718 | int m; |
Stephen Hemminger | c1e20f7 | 2008-07-18 23:02:15 -0700 | [diff] [blame] | 719 | unsigned long rtt; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 720 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 721 | if (icsk->icsk_backoff || !tp->srtt) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 722 | /* This session failed to estimate rtt. Why? |
| 723 | * Probably, no packets returned in time. |
| 724 | * Reset our results. |
| 725 | */ |
| 726 | if (!(dst_metric_locked(dst, RTAX_RTT))) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 727 | dst->metrics[RTAX_RTT - 1] = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 728 | return; |
| 729 | } |
| 730 | |
Stephen Hemminger | c1e20f7 | 2008-07-18 23:02:15 -0700 | [diff] [blame] | 731 | rtt = dst_metric_rtt(dst, RTAX_RTT); |
| 732 | m = rtt - tp->srtt; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 733 | |
| 734 | /* If newly calculated rtt larger than stored one, |
| 735 | * store new one. Otherwise, use EWMA. Remember, |
| 736 | * rtt overestimation is always better than underestimation. |
| 737 | */ |
| 738 | if (!(dst_metric_locked(dst, RTAX_RTT))) { |
| 739 | if (m <= 0) |
Stephen Hemminger | c1e20f7 | 2008-07-18 23:02:15 -0700 | [diff] [blame] | 740 | set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 741 | else |
Stephen Hemminger | c1e20f7 | 2008-07-18 23:02:15 -0700 | [diff] [blame] | 742 | set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 743 | } |
| 744 | |
| 745 | if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { |
Stephen Hemminger | c1e20f7 | 2008-07-18 23:02:15 -0700 | [diff] [blame] | 746 | unsigned long var; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 747 | if (m < 0) |
| 748 | m = -m; |
| 749 | |
| 750 | /* Scale deviation to rttvar fixed point */ |
| 751 | m >>= 1; |
| 752 | if (m < tp->mdev) |
| 753 | m = tp->mdev; |
| 754 | |
Stephen Hemminger | c1e20f7 | 2008-07-18 23:02:15 -0700 | [diff] [blame] | 755 | var = dst_metric_rtt(dst, RTAX_RTTVAR); |
| 756 | if (m >= var) |
| 757 | var = m; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 758 | else |
Stephen Hemminger | c1e20f7 | 2008-07-18 23:02:15 -0700 | [diff] [blame] | 759 | var -= (var - m) >> 2; |
| 760 | |
| 761 | set_dst_metric_rtt(dst, RTAX_RTTVAR, var); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 762 | } |
| 763 | |
Ilpo Järvinen | 0b6a05c | 2009-09-15 01:30:10 -0700 | [diff] [blame] | 764 | if (tcp_in_initial_slowstart(tp)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 765 | /* Slow start still did not finish. */ |
| 766 | if (dst_metric(dst, RTAX_SSTHRESH) && |
| 767 | !dst_metric_locked(dst, RTAX_SSTHRESH) && |
| 768 | (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) |
| 769 | dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1; |
| 770 | if (!dst_metric_locked(dst, RTAX_CWND) && |
| 771 | tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 772 | dst->metrics[RTAX_CWND - 1] = tp->snd_cwnd; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 773 | } else if (tp->snd_cwnd > tp->snd_ssthresh && |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 774 | icsk->icsk_ca_state == TCP_CA_Open) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 775 | /* Cong. avoidance phase, cwnd is reliable. */ |
| 776 | if (!dst_metric_locked(dst, RTAX_SSTHRESH)) |
| 777 | dst->metrics[RTAX_SSTHRESH-1] = |
| 778 | max(tp->snd_cwnd >> 1, tp->snd_ssthresh); |
| 779 | if (!dst_metric_locked(dst, RTAX_CWND)) |
Satoru SATOH | 5ffc02a | 2008-05-04 22:14:42 -0700 | [diff] [blame] | 780 | dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_cwnd) >> 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 781 | } else { |
| 782 | /* Else slow start did not finish, cwnd is non-sense, |
| 783 | ssthresh may be also invalid. |
| 784 | */ |
| 785 | if (!dst_metric_locked(dst, RTAX_CWND)) |
Satoru SATOH | 5ffc02a | 2008-05-04 22:14:42 -0700 | [diff] [blame] | 786 | dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_ssthresh) >> 1; |
| 787 | if (dst_metric(dst, RTAX_SSTHRESH) && |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 788 | !dst_metric_locked(dst, RTAX_SSTHRESH) && |
Satoru SATOH | 5ffc02a | 2008-05-04 22:14:42 -0700 | [diff] [blame] | 789 | tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 790 | dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh; |
| 791 | } |
| 792 | |
| 793 | if (!dst_metric_locked(dst, RTAX_REORDERING)) { |
Satoru SATOH | 5ffc02a | 2008-05-04 22:14:42 -0700 | [diff] [blame] | 794 | if (dst_metric(dst, RTAX_REORDERING) < tp->reordering && |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 795 | tp->reordering != sysctl_tcp_reordering) |
| 796 | dst->metrics[RTAX_REORDERING-1] = tp->reordering; |
| 797 | } |
| 798 | } |
| 799 | } |
| 800 | |
Gerrit Renker | 410e27a | 2008-09-09 13:27:22 +0200 | [diff] [blame] | 801 | /* Numbers are taken from RFC3390. |
| 802 | * |
| 803 | * John Heffner states: |
| 804 | * |
| 805 | * The RFC specifies a window of no more than 4380 bytes |
| 806 | * unless 2*MSS > 4380. Reading the pseudocode in the RFC |
| 807 | * is a bit misleading because they use a clamp at 4380 bytes |
| 808 | * rather than use a multiplier in the relevant range. |
| 809 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 810 | __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst) |
| 811 | { |
| 812 | __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); |
| 813 | |
Gerrit Renker | 410e27a | 2008-09-09 13:27:22 +0200 | [diff] [blame] | 814 | if (!cwnd) { |
| 815 | if (tp->mss_cache > 1460) |
| 816 | cwnd = 2; |
| 817 | else |
| 818 | cwnd = (tp->mss_cache > 1095) ? 3 : 4; |
| 819 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 820 | return min_t(__u32, cwnd, tp->snd_cwnd_clamp); |
| 821 | } |
| 822 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 823 | /* Set slow start threshold and cwnd not falling to slow start */ |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 824 | void tcp_enter_cwr(struct sock *sk, const int set_ssthresh) |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 825 | { |
| 826 | struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 827 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 828 | |
| 829 | tp->prior_ssthresh = 0; |
| 830 | tp->bytes_acked = 0; |
Ilpo Järvinen | e01f9d7 | 2007-03-02 13:27:25 -0800 | [diff] [blame] | 831 | if (icsk->icsk_ca_state < TCP_CA_CWR) { |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 832 | tp->undo_marker = 0; |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 833 | if (set_ssthresh) |
| 834 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 835 | tp->snd_cwnd = min(tp->snd_cwnd, |
| 836 | tcp_packets_in_flight(tp) + 1U); |
| 837 | tp->snd_cwnd_cnt = 0; |
| 838 | tp->high_seq = tp->snd_nxt; |
| 839 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 840 | TCP_ECN_queue_cwr(tp); |
| 841 | |
| 842 | tcp_set_ca_state(sk, TCP_CA_CWR); |
| 843 | } |
| 844 | } |
| 845 | |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 846 | /* |
| 847 | * Packet counting of FACK is based on in-order assumptions, therefore TCP |
| 848 | * disables it when reordering is detected |
| 849 | */ |
| 850 | static void tcp_disable_fack(struct tcp_sock *tp) |
| 851 | { |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 852 | /* RFC3517 uses different metric in lost marker => reset on change */ |
| 853 | if (tcp_is_fack(tp)) |
| 854 | tp->lost_skb_hint = NULL; |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 855 | tp->rx_opt.sack_ok &= ~2; |
| 856 | } |
| 857 | |
Ryousei Takano | 564262c1 | 2007-10-25 23:03:52 -0700 | [diff] [blame] | 858 | /* Take a notice that peer is sending D-SACKs */ |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 859 | static void tcp_dsack_seen(struct tcp_sock *tp) |
| 860 | { |
| 861 | tp->rx_opt.sack_ok |= 4; |
| 862 | } |
| 863 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 864 | /* Initialize metrics on socket. */ |
| 865 | |
| 866 | static void tcp_init_metrics(struct sock *sk) |
| 867 | { |
| 868 | struct tcp_sock *tp = tcp_sk(sk); |
| 869 | struct dst_entry *dst = __sk_dst_get(sk); |
| 870 | |
| 871 | if (dst == NULL) |
| 872 | goto reset; |
| 873 | |
| 874 | dst_confirm(dst); |
| 875 | |
| 876 | if (dst_metric_locked(dst, RTAX_CWND)) |
| 877 | tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND); |
| 878 | if (dst_metric(dst, RTAX_SSTHRESH)) { |
| 879 | tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH); |
| 880 | if (tp->snd_ssthresh > tp->snd_cwnd_clamp) |
| 881 | tp->snd_ssthresh = tp->snd_cwnd_clamp; |
| 882 | } |
| 883 | if (dst_metric(dst, RTAX_REORDERING) && |
| 884 | tp->reordering != dst_metric(dst, RTAX_REORDERING)) { |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 885 | tcp_disable_fack(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 886 | tp->reordering = dst_metric(dst, RTAX_REORDERING); |
| 887 | } |
| 888 | |
| 889 | if (dst_metric(dst, RTAX_RTT) == 0) |
| 890 | goto reset; |
| 891 | |
Stephen Hemminger | c1e20f7 | 2008-07-18 23:02:15 -0700 | [diff] [blame] | 892 | if (!tp->srtt && dst_metric_rtt(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 893 | goto reset; |
| 894 | |
| 895 | /* Initial rtt is determined from SYN,SYN-ACK. |
| 896 | * The segment is small and rtt may appear much |
| 897 | * less than real one. Use per-dst memory |
| 898 | * to make it more realistic. |
| 899 | * |
| 900 | * A bit of theory. RTT is time passed after "normal" sized packet |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 901 | * is sent until it is ACKed. In normal circumstances sending small |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 902 | * packets force peer to delay ACKs and calculation is correct too. |
| 903 | * The algorithm is adaptive and, provided we follow specs, it |
| 904 | * NEVER underestimate RTT. BUT! If peer tries to make some clever |
| 905 | * tricks sort of "quick acks" for time long enough to decrease RTT |
| 906 | * to low value, and then abruptly stops to do it and starts to delay |
| 907 | * ACKs, wait for troubles. |
| 908 | */ |
Stephen Hemminger | c1e20f7 | 2008-07-18 23:02:15 -0700 | [diff] [blame] | 909 | if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) { |
| 910 | tp->srtt = dst_metric_rtt(dst, RTAX_RTT); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 911 | tp->rtt_seq = tp->snd_nxt; |
| 912 | } |
Stephen Hemminger | c1e20f7 | 2008-07-18 23:02:15 -0700 | [diff] [blame] | 913 | if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) { |
| 914 | tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR); |
Satoru SATOH | 488faa2 | 2007-12-16 14:00:19 -0800 | [diff] [blame] | 915 | tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 916 | } |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 917 | tcp_set_rto(sk); |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 918 | if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 919 | goto reset; |
Ilpo Järvinen | 86bceba | 2009-04-14 02:08:53 -0700 | [diff] [blame] | 920 | |
| 921 | cwnd: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 922 | tp->snd_cwnd = tcp_init_cwnd(tp, dst); |
| 923 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 924 | return; |
| 925 | |
| 926 | reset: |
| 927 | /* Play conservative. If timestamps are not |
| 928 | * supported, TCP will fail to recalculate correct |
| 929 | * rtt, if initial rto is too small. FORGET ALL AND RESET! |
| 930 | */ |
| 931 | if (!tp->rx_opt.saw_tstamp && tp->srtt) { |
| 932 | tp->srtt = 0; |
| 933 | tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT; |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 934 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 935 | } |
Ilpo Järvinen | 86bceba | 2009-04-14 02:08:53 -0700 | [diff] [blame] | 936 | goto cwnd; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 937 | } |
| 938 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 939 | static void tcp_update_reordering(struct sock *sk, const int metric, |
| 940 | const int ts) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 941 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 942 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 943 | if (metric > tp->reordering) { |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 944 | int mib_idx; |
| 945 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 946 | tp->reordering = min(TCP_MAX_REORDERING, metric); |
| 947 | |
| 948 | /* This exciting event is worth to be remembered. 8) */ |
| 949 | if (ts) |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 950 | mib_idx = LINUX_MIB_TCPTSREORDER; |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 951 | else if (tcp_is_reno(tp)) |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 952 | mib_idx = LINUX_MIB_TCPRENOREORDER; |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 953 | else if (tcp_is_fack(tp)) |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 954 | mib_idx = LINUX_MIB_TCPFACKREORDER; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 955 | else |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 956 | mib_idx = LINUX_MIB_TCPSACKREORDER; |
| 957 | |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 958 | NET_INC_STATS_BH(sock_net(sk), mib_idx); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 959 | #if FASTRETRANS_DEBUG > 1 |
| 960 | printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 961 | tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 962 | tp->reordering, |
| 963 | tp->fackets_out, |
| 964 | tp->sacked_out, |
| 965 | tp->undo_marker ? tp->undo_retrans : 0); |
| 966 | #endif |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 967 | tcp_disable_fack(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 968 | } |
| 969 | } |
| 970 | |
Ilpo Järvinen | 006f582 | 2008-09-20 21:20:20 -0700 | [diff] [blame] | 971 | /* This must be called before lost_out is incremented */ |
Ilpo Järvinen | c8c213f | 2008-09-20 21:18:55 -0700 | [diff] [blame] | 972 | static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) |
| 973 | { |
Ilpo Järvinen | 006f582 | 2008-09-20 21:20:20 -0700 | [diff] [blame] | 974 | if ((tp->retransmit_skb_hint == NULL) || |
Ilpo Järvinen | c8c213f | 2008-09-20 21:18:55 -0700 | [diff] [blame] | 975 | before(TCP_SKB_CB(skb)->seq, |
| 976 | TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) |
Ilpo Järvinen | 006f582 | 2008-09-20 21:20:20 -0700 | [diff] [blame] | 977 | tp->retransmit_skb_hint = skb; |
| 978 | |
| 979 | if (!tp->lost_out || |
| 980 | after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high)) |
| 981 | tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; |
Ilpo Järvinen | c8c213f | 2008-09-20 21:18:55 -0700 | [diff] [blame] | 982 | } |
| 983 | |
Ilpo Järvinen | 41ea36e | 2008-09-20 21:19:22 -0700 | [diff] [blame] | 984 | static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb) |
| 985 | { |
| 986 | if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { |
| 987 | tcp_verify_retransmit_hint(tp, skb); |
| 988 | |
| 989 | tp->lost_out += tcp_skb_pcount(skb); |
| 990 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; |
| 991 | } |
| 992 | } |
| 993 | |
Ilpo Järvinen | e1aa680 | 2008-11-24 21:11:55 -0800 | [diff] [blame] | 994 | static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, |
| 995 | struct sk_buff *skb) |
Ilpo Järvinen | 006f582 | 2008-09-20 21:20:20 -0700 | [diff] [blame] | 996 | { |
| 997 | tcp_verify_retransmit_hint(tp, skb); |
| 998 | |
| 999 | if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { |
| 1000 | tp->lost_out += tcp_skb_pcount(skb); |
| 1001 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; |
| 1002 | } |
| 1003 | } |
| 1004 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1005 | /* This procedure tags the retransmission queue when SACKs arrive. |
| 1006 | * |
| 1007 | * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). |
| 1008 | * Packets in queue with these bits set are counted in variables |
| 1009 | * sacked_out, retrans_out and lost_out, correspondingly. |
| 1010 | * |
| 1011 | * Valid combinations are: |
| 1012 | * Tag InFlight Description |
| 1013 | * 0 1 - orig segment is in flight. |
| 1014 | * S 0 - nothing flies, orig reached receiver. |
| 1015 | * L 0 - nothing flies, orig lost by net. |
| 1016 | * R 2 - both orig and retransmit are in flight. |
| 1017 | * L|R 1 - orig is lost, retransmit is in flight. |
| 1018 | * S|R 1 - orig reached receiver, retrans is still in flight. |
| 1019 | * (L|S|R is logically valid, it could occur when L|R is sacked, |
| 1020 | * but it is equivalent to plain S and code short-curcuits it to S. |
| 1021 | * L|S is logically invalid, it would mean -1 packet in flight 8)) |
| 1022 | * |
| 1023 | * These 6 states form finite state machine, controlled by the following events: |
| 1024 | * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) |
| 1025 | * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) |
| 1026 | * 3. Loss detection event of one of three flavors: |
| 1027 | * A. Scoreboard estimator decided the packet is lost. |
| 1028 | * A'. Reno "three dupacks" marks head of queue lost. |
| 1029 | * A''. Its FACK modfication, head until snd.fack is lost. |
| 1030 | * B. SACK arrives sacking data transmitted after never retransmitted |
| 1031 | * hole was sent out. |
| 1032 | * C. SACK arrives sacking SND.NXT at the moment, when the |
| 1033 | * segment was retransmitted. |
| 1034 | * 4. D-SACK added new rule: D-SACK changes any tag to S. |
| 1035 | * |
| 1036 | * It is pleasant to note, that state diagram turns out to be commutative, |
| 1037 | * so that we are allowed not to be bothered by order of our actions, |
| 1038 | * when multiple events arrive simultaneously. (see the function below). |
| 1039 | * |
| 1040 | * Reordering detection. |
| 1041 | * -------------------- |
| 1042 | * Reordering metric is maximal distance, which a packet can be displaced |
| 1043 | * in packet stream. With SACKs we can estimate it: |
| 1044 | * |
| 1045 | * 1. SACK fills old hole and the corresponding segment was not |
| 1046 | * ever retransmitted -> reordering. Alas, we cannot use it |
| 1047 | * when segment was retransmitted. |
| 1048 | * 2. The last flaw is solved with D-SACK. D-SACK arrives |
| 1049 | * for retransmitted and already SACKed segment -> reordering.. |
| 1050 | * Both of these heuristics are not used in Loss state, when we cannot |
| 1051 | * account for retransmits accurately. |
Ilpo Järvinen | 5b3c988 | 2007-08-24 22:54:44 -0700 | [diff] [blame] | 1052 | * |
| 1053 | * SACK block validation. |
| 1054 | * ---------------------- |
| 1055 | * |
| 1056 | * SACK block range validation checks that the received SACK block fits to |
| 1057 | * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. |
| 1058 | * Note that SND.UNA is not included to the range though being valid because |
Ilpo Järvinen | 0e83533 | 2007-10-01 15:28:17 -0700 | [diff] [blame] | 1059 | * it means that the receiver is rather inconsistent with itself reporting |
| 1060 | * SACK reneging when it should advance SND.UNA. Such SACK block this is |
| 1061 | * perfectly valid, however, in light of RFC2018 which explicitly states |
| 1062 | * that "SACK block MUST reflect the newest segment. Even if the newest |
| 1063 | * segment is going to be discarded ...", not that it looks very clever |
| 1064 | * in case of head skb. Due to potentional receiver driven attacks, we |
| 1065 | * choose to avoid immediate execution of a walk in write queue due to |
| 1066 | * reneging and defer head skb's loss recovery to standard loss recovery |
| 1067 | * procedure that will eventually trigger (nothing forbids us doing this). |
Ilpo Järvinen | 5b3c988 | 2007-08-24 22:54:44 -0700 | [diff] [blame] | 1068 | * |
| 1069 | * Implements also blockage to start_seq wrap-around. Problem lies in the |
| 1070 | * fact that though start_seq (s) is before end_seq (i.e., not reversed), |
| 1071 | * there's no guarantee that it will be before snd_nxt (n). The problem |
| 1072 | * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt |
| 1073 | * wrap (s_w): |
| 1074 | * |
| 1075 | * <- outs wnd -> <- wrapzone -> |
| 1076 | * u e n u_w e_w s n_w |
| 1077 | * | | | | | | | |
| 1078 | * |<------------+------+----- TCP seqno space --------------+---------->| |
| 1079 | * ...-- <2^31 ->| |<--------... |
| 1080 | * ...---- >2^31 ------>| |<--------... |
| 1081 | * |
| 1082 | * Current code wouldn't be vulnerable but it's better still to discard such |
| 1083 | * crazy SACK blocks. Doing this check for start_seq alone closes somewhat |
| 1084 | * similar case (end_seq after snd_nxt wrap) as earlier reversed check in |
| 1085 | * snd_nxt wrap -> snd_una region will then become "well defined", i.e., |
| 1086 | * equal to the ideal case (infinite seqno space without wrap caused issues). |
| 1087 | * |
| 1088 | * With D-SACK the lower bound is extended to cover sequence space below |
| 1089 | * SND.UNA down to undo_marker, which is the last point of interest. Yet |
Ryousei Takano | 564262c1 | 2007-10-25 23:03:52 -0700 | [diff] [blame] | 1090 | * again, D-SACK block must not to go across snd_una (for the same reason as |
Ilpo Järvinen | 5b3c988 | 2007-08-24 22:54:44 -0700 | [diff] [blame] | 1091 | * for the normal SACK blocks, explained above). But there all simplicity |
| 1092 | * ends, TCP might receive valid D-SACKs below that. As long as they reside |
| 1093 | * fully below undo_marker they do not affect behavior in anyway and can |
| 1094 | * therefore be safely ignored. In rare cases (which are more or less |
| 1095 | * theoretical ones), the D-SACK will nicely cross that boundary due to skb |
| 1096 | * fragmentation and packet reordering past skb's retransmission. To consider |
| 1097 | * them correctly, the acceptable range must be extended even more though |
| 1098 | * the exact amount is rather hard to quantify. However, tp->max_window can |
| 1099 | * be used as an exaggerated estimate. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1100 | */ |
Ilpo Järvinen | 5b3c988 | 2007-08-24 22:54:44 -0700 | [diff] [blame] | 1101 | static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack, |
| 1102 | u32 start_seq, u32 end_seq) |
| 1103 | { |
| 1104 | /* Too far in future, or reversed (interpretation is ambiguous) */ |
| 1105 | if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) |
| 1106 | return 0; |
| 1107 | |
| 1108 | /* Nasty start_seq wrap-around check (see comments above) */ |
| 1109 | if (!before(start_seq, tp->snd_nxt)) |
| 1110 | return 0; |
| 1111 | |
Ryousei Takano | 564262c1 | 2007-10-25 23:03:52 -0700 | [diff] [blame] | 1112 | /* In outstanding window? ...This is valid exit for D-SACKs too. |
Ilpo Järvinen | 5b3c988 | 2007-08-24 22:54:44 -0700 | [diff] [blame] | 1113 | * start_seq == snd_una is non-sensical (see comments above) |
| 1114 | */ |
| 1115 | if (after(start_seq, tp->snd_una)) |
| 1116 | return 1; |
| 1117 | |
| 1118 | if (!is_dsack || !tp->undo_marker) |
| 1119 | return 0; |
| 1120 | |
| 1121 | /* ...Then it's D-SACK, and must reside below snd_una completely */ |
| 1122 | if (!after(end_seq, tp->snd_una)) |
| 1123 | return 0; |
| 1124 | |
| 1125 | if (!before(start_seq, tp->undo_marker)) |
| 1126 | return 1; |
| 1127 | |
| 1128 | /* Too old */ |
| 1129 | if (!after(end_seq, tp->undo_marker)) |
| 1130 | return 0; |
| 1131 | |
| 1132 | /* Undo_marker boundary crossing (overestimates a lot). Known already: |
| 1133 | * start_seq < undo_marker and end_seq >= undo_marker. |
| 1134 | */ |
| 1135 | return !before(start_seq, end_seq - tp->max_window); |
| 1136 | } |
| 1137 | |
Ilpo Järvinen | 1c1e87e | 2007-10-10 02:45:32 -0700 | [diff] [blame] | 1138 | /* Check for lost retransmit. This superb idea is borrowed from "ratehalving". |
| 1139 | * Event "C". Later note: FACK people cheated me again 8), we have to account |
| 1140 | * for reordering! Ugly, but should help. |
Ilpo Järvinen | f785a8e | 2007-10-11 17:35:41 -0700 | [diff] [blame] | 1141 | * |
| 1142 | * Search retransmitted skbs from write_queue that were sent when snd_nxt was |
| 1143 | * less than what is now known to be received by the other end (derived from |
Ilpo Järvinen | 9f58f3b | 2007-11-15 19:42:54 -0800 | [diff] [blame] | 1144 | * highest SACK block). Also calculate the lowest snd_nxt among the remaining |
| 1145 | * retransmitted skbs to avoid some costly processing per ACKs. |
Ilpo Järvinen | 1c1e87e | 2007-10-10 02:45:32 -0700 | [diff] [blame] | 1146 | */ |
Ilpo Järvinen | 407ef1d | 2007-12-02 00:47:57 +0200 | [diff] [blame] | 1147 | static void tcp_mark_lost_retrans(struct sock *sk) |
Ilpo Järvinen | 1c1e87e | 2007-10-10 02:45:32 -0700 | [diff] [blame] | 1148 | { |
Ilpo Järvinen | 9f58f3b | 2007-11-15 19:42:54 -0800 | [diff] [blame] | 1149 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Ilpo Järvinen | 1c1e87e | 2007-10-10 02:45:32 -0700 | [diff] [blame] | 1150 | struct tcp_sock *tp = tcp_sk(sk); |
| 1151 | struct sk_buff *skb; |
Ilpo Järvinen | f785a8e | 2007-10-11 17:35:41 -0700 | [diff] [blame] | 1152 | int cnt = 0; |
Ilpo Järvinen | df2e014 | 2007-10-18 05:07:57 -0700 | [diff] [blame] | 1153 | u32 new_low_seq = tp->snd_nxt; |
Ilpo Järvinen | 6859d49 | 2007-12-02 00:48:06 +0200 | [diff] [blame] | 1154 | u32 received_upto = tcp_highest_sack_seq(tp); |
Ilpo Järvinen | 9f58f3b | 2007-11-15 19:42:54 -0800 | [diff] [blame] | 1155 | |
| 1156 | if (!tcp_is_fack(tp) || !tp->retrans_out || |
| 1157 | !after(received_upto, tp->lost_retrans_low) || |
| 1158 | icsk->icsk_ca_state != TCP_CA_Recovery) |
Ilpo Järvinen | 407ef1d | 2007-12-02 00:47:57 +0200 | [diff] [blame] | 1159 | return; |
Ilpo Järvinen | 1c1e87e | 2007-10-10 02:45:32 -0700 | [diff] [blame] | 1160 | |
| 1161 | tcp_for_write_queue(skb, sk) { |
| 1162 | u32 ack_seq = TCP_SKB_CB(skb)->ack_seq; |
| 1163 | |
| 1164 | if (skb == tcp_send_head(sk)) |
| 1165 | break; |
Ilpo Järvinen | f785a8e | 2007-10-11 17:35:41 -0700 | [diff] [blame] | 1166 | if (cnt == tp->retrans_out) |
Ilpo Järvinen | 1c1e87e | 2007-10-10 02:45:32 -0700 | [diff] [blame] | 1167 | break; |
| 1168 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) |
| 1169 | continue; |
| 1170 | |
Ilpo Järvinen | f785a8e | 2007-10-11 17:35:41 -0700 | [diff] [blame] | 1171 | if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) |
| 1172 | continue; |
| 1173 | |
Ilpo Järvinen | d0af416 | 2009-02-28 04:44:32 +0000 | [diff] [blame] | 1174 | /* TODO: We would like to get rid of tcp_is_fack(tp) only |
| 1175 | * constraint here (see above) but figuring out that at |
| 1176 | * least tp->reordering SACK blocks reside between ack_seq |
| 1177 | * and received_upto is not easy task to do cheaply with |
| 1178 | * the available datastructures. |
| 1179 | * |
| 1180 | * Whether FACK should check here for tp->reordering segs |
| 1181 | * in-between one could argue for either way (it would be |
| 1182 | * rather simple to implement as we could count fack_count |
| 1183 | * during the walk and do tp->fackets_out - fack_count). |
| 1184 | */ |
| 1185 | if (after(received_upto, ack_seq)) { |
Ilpo Järvinen | 1c1e87e | 2007-10-10 02:45:32 -0700 | [diff] [blame] | 1186 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; |
| 1187 | tp->retrans_out -= tcp_skb_pcount(skb); |
| 1188 | |
Ilpo Järvinen | 006f582 | 2008-09-20 21:20:20 -0700 | [diff] [blame] | 1189 | tcp_skb_mark_lost_uncond_verify(tp, skb); |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 1190 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT); |
Ilpo Järvinen | f785a8e | 2007-10-11 17:35:41 -0700 | [diff] [blame] | 1191 | } else { |
Ilpo Järvinen | df2e014 | 2007-10-18 05:07:57 -0700 | [diff] [blame] | 1192 | if (before(ack_seq, new_low_seq)) |
Ilpo Järvinen | b08d6cb | 2007-10-11 17:36:13 -0700 | [diff] [blame] | 1193 | new_low_seq = ack_seq; |
Ilpo Järvinen | f785a8e | 2007-10-11 17:35:41 -0700 | [diff] [blame] | 1194 | cnt += tcp_skb_pcount(skb); |
Ilpo Järvinen | 1c1e87e | 2007-10-10 02:45:32 -0700 | [diff] [blame] | 1195 | } |
| 1196 | } |
Ilpo Järvinen | b08d6cb | 2007-10-11 17:36:13 -0700 | [diff] [blame] | 1197 | |
| 1198 | if (tp->retrans_out) |
| 1199 | tp->lost_retrans_low = new_low_seq; |
Ilpo Järvinen | 1c1e87e | 2007-10-10 02:45:32 -0700 | [diff] [blame] | 1200 | } |
Ilpo Järvinen | 5b3c988 | 2007-08-24 22:54:44 -0700 | [diff] [blame] | 1201 | |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 1202 | static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb, |
David S. Miller | d06e021 | 2007-06-18 22:43:06 -0700 | [diff] [blame] | 1203 | struct tcp_sack_block_wire *sp, int num_sacks, |
| 1204 | u32 prior_snd_una) |
| 1205 | { |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 1206 | struct tcp_sock *tp = tcp_sk(sk); |
Harvey Harrison | d3e2ce3 | 2008-05-02 16:26:16 -0700 | [diff] [blame] | 1207 | u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); |
| 1208 | u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); |
David S. Miller | d06e021 | 2007-06-18 22:43:06 -0700 | [diff] [blame] | 1209 | int dup_sack = 0; |
| 1210 | |
| 1211 | if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { |
| 1212 | dup_sack = 1; |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 1213 | tcp_dsack_seen(tp); |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 1214 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV); |
David S. Miller | d06e021 | 2007-06-18 22:43:06 -0700 | [diff] [blame] | 1215 | } else if (num_sacks > 1) { |
Harvey Harrison | d3e2ce3 | 2008-05-02 16:26:16 -0700 | [diff] [blame] | 1216 | u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); |
| 1217 | u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); |
David S. Miller | d06e021 | 2007-06-18 22:43:06 -0700 | [diff] [blame] | 1218 | |
| 1219 | if (!after(end_seq_0, end_seq_1) && |
| 1220 | !before(start_seq_0, start_seq_1)) { |
| 1221 | dup_sack = 1; |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 1222 | tcp_dsack_seen(tp); |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 1223 | NET_INC_STATS_BH(sock_net(sk), |
| 1224 | LINUX_MIB_TCPDSACKOFORECV); |
David S. Miller | d06e021 | 2007-06-18 22:43:06 -0700 | [diff] [blame] | 1225 | } |
| 1226 | } |
| 1227 | |
| 1228 | /* D-SACK for already forgotten data... Do dumb counting. */ |
| 1229 | if (dup_sack && |
| 1230 | !after(end_seq_0, prior_snd_una) && |
| 1231 | after(end_seq_0, tp->undo_marker)) |
| 1232 | tp->undo_retrans--; |
| 1233 | |
| 1234 | return dup_sack; |
| 1235 | } |
| 1236 | |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1237 | struct tcp_sacktag_state { |
| 1238 | int reord; |
| 1239 | int fack_count; |
| 1240 | int flag; |
| 1241 | }; |
| 1242 | |
Ilpo Järvinen | d193594 | 2007-10-11 17:34:25 -0700 | [diff] [blame] | 1243 | /* Check if skb is fully within the SACK block. In presence of GSO skbs, |
| 1244 | * the incoming SACK may not exactly match but we can find smaller MSS |
| 1245 | * aligned portion of it that matches. Therefore we might need to fragment |
| 1246 | * which may fail and creates some hassle (caller must handle error case |
| 1247 | * returns). |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1248 | * |
| 1249 | * FIXME: this could be merged to shift decision code |
Ilpo Järvinen | d193594 | 2007-10-11 17:34:25 -0700 | [diff] [blame] | 1250 | */ |
Adrian Bunk | 0f79efd | 2007-10-26 03:57:36 -0700 | [diff] [blame] | 1251 | static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, |
| 1252 | u32 start_seq, u32 end_seq) |
Ilpo Järvinen | d193594 | 2007-10-11 17:34:25 -0700 | [diff] [blame] | 1253 | { |
| 1254 | int in_sack, err; |
| 1255 | unsigned int pkt_len; |
Ilpo Järvinen | adb92db | 2008-11-24 21:13:50 -0800 | [diff] [blame] | 1256 | unsigned int mss; |
Ilpo Järvinen | d193594 | 2007-10-11 17:34:25 -0700 | [diff] [blame] | 1257 | |
| 1258 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && |
| 1259 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); |
| 1260 | |
| 1261 | if (tcp_skb_pcount(skb) > 1 && !in_sack && |
| 1262 | after(TCP_SKB_CB(skb)->end_seq, start_seq)) { |
Ilpo Järvinen | adb92db | 2008-11-24 21:13:50 -0800 | [diff] [blame] | 1263 | mss = tcp_skb_mss(skb); |
Ilpo Järvinen | d193594 | 2007-10-11 17:34:25 -0700 | [diff] [blame] | 1264 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); |
| 1265 | |
Ilpo Järvinen | adb92db | 2008-11-24 21:13:50 -0800 | [diff] [blame] | 1266 | if (!in_sack) { |
Ilpo Järvinen | d193594 | 2007-10-11 17:34:25 -0700 | [diff] [blame] | 1267 | pkt_len = start_seq - TCP_SKB_CB(skb)->seq; |
Ilpo Järvinen | adb92db | 2008-11-24 21:13:50 -0800 | [diff] [blame] | 1268 | if (pkt_len < mss) |
| 1269 | pkt_len = mss; |
| 1270 | } else { |
Ilpo Järvinen | d193594 | 2007-10-11 17:34:25 -0700 | [diff] [blame] | 1271 | pkt_len = end_seq - TCP_SKB_CB(skb)->seq; |
Ilpo Järvinen | adb92db | 2008-11-24 21:13:50 -0800 | [diff] [blame] | 1272 | if (pkt_len < mss) |
| 1273 | return -EINVAL; |
| 1274 | } |
| 1275 | |
| 1276 | /* Round if necessary so that SACKs cover only full MSSes |
| 1277 | * and/or the remaining small portion (if present) |
| 1278 | */ |
| 1279 | if (pkt_len > mss) { |
| 1280 | unsigned int new_len = (pkt_len / mss) * mss; |
| 1281 | if (!in_sack && new_len < pkt_len) { |
| 1282 | new_len += mss; |
| 1283 | if (new_len > skb->len) |
| 1284 | return 0; |
| 1285 | } |
| 1286 | pkt_len = new_len; |
| 1287 | } |
| 1288 | err = tcp_fragment(sk, skb, pkt_len, mss); |
Ilpo Järvinen | d193594 | 2007-10-11 17:34:25 -0700 | [diff] [blame] | 1289 | if (err < 0) |
| 1290 | return err; |
| 1291 | } |
| 1292 | |
| 1293 | return in_sack; |
| 1294 | } |
| 1295 | |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1296 | static u8 tcp_sacktag_one(struct sk_buff *skb, struct sock *sk, |
| 1297 | struct tcp_sacktag_state *state, |
| 1298 | int dup_sack, int pcount) |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1299 | { |
Ilpo Järvinen | 6859d49 | 2007-12-02 00:48:06 +0200 | [diff] [blame] | 1300 | struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1301 | u8 sacked = TCP_SKB_CB(skb)->sacked; |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1302 | int fack_count = state->fack_count; |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1303 | |
| 1304 | /* Account D-SACK for retransmitted packet. */ |
| 1305 | if (dup_sack && (sacked & TCPCB_RETRANS)) { |
| 1306 | if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker)) |
| 1307 | tp->undo_retrans--; |
Ilpo Järvinen | ede9f3b | 2007-12-02 00:47:58 +0200 | [diff] [blame] | 1308 | if (sacked & TCPCB_SACKED_ACKED) |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1309 | state->reord = min(fack_count, state->reord); |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1310 | } |
| 1311 | |
| 1312 | /* Nothing to do; acked frame is about to be dropped (was ACKed). */ |
| 1313 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1314 | return sacked; |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1315 | |
| 1316 | if (!(sacked & TCPCB_SACKED_ACKED)) { |
| 1317 | if (sacked & TCPCB_SACKED_RETRANS) { |
| 1318 | /* If the segment is not tagged as lost, |
| 1319 | * we do not clear RETRANS, believing |
| 1320 | * that retransmission is still in flight. |
| 1321 | */ |
| 1322 | if (sacked & TCPCB_LOST) { |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1323 | sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); |
Ilpo Järvinen | f58b22f | 2008-11-24 21:14:43 -0800 | [diff] [blame] | 1324 | tp->lost_out -= pcount; |
| 1325 | tp->retrans_out -= pcount; |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1326 | } |
| 1327 | } else { |
| 1328 | if (!(sacked & TCPCB_RETRANS)) { |
| 1329 | /* New sack for not retransmitted frame, |
| 1330 | * which was in hole. It is reordering. |
| 1331 | */ |
| 1332 | if (before(TCP_SKB_CB(skb)->seq, |
| 1333 | tcp_highest_sack_seq(tp))) |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1334 | state->reord = min(fack_count, |
| 1335 | state->reord); |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1336 | |
| 1337 | /* SACK enhanced F-RTO (RFC4138; Appendix B) */ |
| 1338 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1339 | state->flag |= FLAG_ONLY_ORIG_SACKED; |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1340 | } |
| 1341 | |
| 1342 | if (sacked & TCPCB_LOST) { |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1343 | sacked &= ~TCPCB_LOST; |
Ilpo Järvinen | f58b22f | 2008-11-24 21:14:43 -0800 | [diff] [blame] | 1344 | tp->lost_out -= pcount; |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1345 | } |
| 1346 | } |
| 1347 | |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1348 | sacked |= TCPCB_SACKED_ACKED; |
| 1349 | state->flag |= FLAG_DATA_SACKED; |
Ilpo Järvinen | f58b22f | 2008-11-24 21:14:43 -0800 | [diff] [blame] | 1350 | tp->sacked_out += pcount; |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1351 | |
Ilpo Järvinen | f58b22f | 2008-11-24 21:14:43 -0800 | [diff] [blame] | 1352 | fack_count += pcount; |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1353 | |
| 1354 | /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ |
| 1355 | if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) && |
| 1356 | before(TCP_SKB_CB(skb)->seq, |
| 1357 | TCP_SKB_CB(tp->lost_skb_hint)->seq)) |
Ilpo Järvinen | f58b22f | 2008-11-24 21:14:43 -0800 | [diff] [blame] | 1358 | tp->lost_cnt_hint += pcount; |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1359 | |
| 1360 | if (fack_count > tp->fackets_out) |
| 1361 | tp->fackets_out = fack_count; |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1362 | } |
| 1363 | |
| 1364 | /* D-SACK. We can detect redundant retransmission in S|R and plain R |
| 1365 | * frames and clear it. undo_retrans is decreased above, L|R frames |
| 1366 | * are accounted above as well. |
| 1367 | */ |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1368 | if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { |
| 1369 | sacked &= ~TCPCB_SACKED_RETRANS; |
Ilpo Järvinen | f58b22f | 2008-11-24 21:14:43 -0800 | [diff] [blame] | 1370 | tp->retrans_out -= pcount; |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1371 | } |
| 1372 | |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1373 | return sacked; |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1374 | } |
| 1375 | |
Ilpo Järvinen | 5013316 | 2008-12-05 22:42:41 -0800 | [diff] [blame] | 1376 | static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb, |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1377 | struct tcp_sacktag_state *state, |
Ilpo Järvinen | 9ec06ff | 2009-03-01 00:21:36 -0800 | [diff] [blame] | 1378 | unsigned int pcount, int shifted, int mss, |
| 1379 | int dup_sack) |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1380 | { |
| 1381 | struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 5013316 | 2008-12-05 22:42:41 -0800 | [diff] [blame] | 1382 | struct sk_buff *prev = tcp_write_queue_prev(sk, skb); |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1383 | |
| 1384 | BUG_ON(!pcount); |
| 1385 | |
Ilpo Järvinen | 92ee76b | 2008-11-24 21:26:56 -0800 | [diff] [blame] | 1386 | /* Tweak before seqno plays */ |
| 1387 | if (!tcp_is_fack(tp) && tcp_is_sack(tp) && tp->lost_skb_hint && |
| 1388 | !before(TCP_SKB_CB(tp->lost_skb_hint)->seq, TCP_SKB_CB(skb)->seq)) |
| 1389 | tp->lost_cnt_hint += pcount; |
| 1390 | |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1391 | TCP_SKB_CB(prev)->end_seq += shifted; |
| 1392 | TCP_SKB_CB(skb)->seq += shifted; |
| 1393 | |
| 1394 | skb_shinfo(prev)->gso_segs += pcount; |
| 1395 | BUG_ON(skb_shinfo(skb)->gso_segs < pcount); |
| 1396 | skb_shinfo(skb)->gso_segs -= pcount; |
| 1397 | |
| 1398 | /* When we're adding to gso_segs == 1, gso_size will be zero, |
| 1399 | * in theory this shouldn't be necessary but as long as DSACK |
| 1400 | * code can come after this skb later on it's better to keep |
| 1401 | * setting gso_size to something. |
| 1402 | */ |
| 1403 | if (!skb_shinfo(prev)->gso_size) { |
| 1404 | skb_shinfo(prev)->gso_size = mss; |
| 1405 | skb_shinfo(prev)->gso_type = sk->sk_gso_type; |
| 1406 | } |
| 1407 | |
| 1408 | /* CHECKME: To clear or not to clear? Mimics normal skb currently */ |
| 1409 | if (skb_shinfo(skb)->gso_segs <= 1) { |
| 1410 | skb_shinfo(skb)->gso_size = 0; |
| 1411 | skb_shinfo(skb)->gso_type = 0; |
| 1412 | } |
| 1413 | |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1414 | /* We discard results */ |
Ilpo Järvinen | 9ec06ff | 2009-03-01 00:21:36 -0800 | [diff] [blame] | 1415 | tcp_sacktag_one(skb, sk, state, dup_sack, pcount); |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1416 | |
| 1417 | /* Difference in this won't matter, both ACKed by the same cumul. ACK */ |
| 1418 | TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); |
| 1419 | |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1420 | if (skb->len > 0) { |
| 1421 | BUG_ON(!tcp_skb_pcount(skb)); |
Ilpo Järvinen | 111cc8b | 2008-11-24 21:27:22 -0800 | [diff] [blame] | 1422 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED); |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1423 | return 0; |
| 1424 | } |
| 1425 | |
| 1426 | /* Whole SKB was eaten :-) */ |
| 1427 | |
Ilpo Järvinen | 92ee76b | 2008-11-24 21:26:56 -0800 | [diff] [blame] | 1428 | if (skb == tp->retransmit_skb_hint) |
| 1429 | tp->retransmit_skb_hint = prev; |
| 1430 | if (skb == tp->scoreboard_skb_hint) |
| 1431 | tp->scoreboard_skb_hint = prev; |
| 1432 | if (skb == tp->lost_skb_hint) { |
| 1433 | tp->lost_skb_hint = prev; |
| 1434 | tp->lost_cnt_hint -= tcp_skb_pcount(prev); |
| 1435 | } |
| 1436 | |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1437 | TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(prev)->flags; |
| 1438 | if (skb == tcp_highest_sack(sk)) |
| 1439 | tcp_advance_highest_sack(sk, skb); |
| 1440 | |
| 1441 | tcp_unlink_write_queue(skb, sk); |
| 1442 | sk_wmem_free_skb(sk, skb); |
| 1443 | |
Ilpo Järvinen | 111cc8b | 2008-11-24 21:27:22 -0800 | [diff] [blame] | 1444 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED); |
| 1445 | |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1446 | return 1; |
| 1447 | } |
| 1448 | |
| 1449 | /* I wish gso_size would have a bit more sane initialization than |
| 1450 | * something-or-zero which complicates things |
| 1451 | */ |
Ilpo Järvinen | 775ffab | 2008-12-05 22:41:26 -0800 | [diff] [blame] | 1452 | static int tcp_skb_seglen(struct sk_buff *skb) |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1453 | { |
Ilpo Järvinen | 775ffab | 2008-12-05 22:41:26 -0800 | [diff] [blame] | 1454 | return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1455 | } |
| 1456 | |
| 1457 | /* Shifting pages past head area doesn't work */ |
| 1458 | static int skb_can_shift(struct sk_buff *skb) |
| 1459 | { |
| 1460 | return !skb_headlen(skb) && skb_is_nonlinear(skb); |
| 1461 | } |
| 1462 | |
| 1463 | /* Try collapsing SACK blocks spanning across multiple skbs to a single |
| 1464 | * skb. |
| 1465 | */ |
| 1466 | static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1467 | struct tcp_sacktag_state *state, |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1468 | u32 start_seq, u32 end_seq, |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1469 | int dup_sack) |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1470 | { |
| 1471 | struct tcp_sock *tp = tcp_sk(sk); |
| 1472 | struct sk_buff *prev; |
| 1473 | int mss; |
| 1474 | int pcount = 0; |
| 1475 | int len; |
| 1476 | int in_sack; |
| 1477 | |
| 1478 | if (!sk_can_gso(sk)) |
| 1479 | goto fallback; |
| 1480 | |
| 1481 | /* Normally R but no L won't result in plain S */ |
| 1482 | if (!dup_sack && |
Ilpo Järvinen | 9969ca5 | 2008-12-05 22:41:06 -0800 | [diff] [blame] | 1483 | (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1484 | goto fallback; |
| 1485 | if (!skb_can_shift(skb)) |
| 1486 | goto fallback; |
| 1487 | /* This frame is about to be dropped (was ACKed). */ |
| 1488 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) |
| 1489 | goto fallback; |
| 1490 | |
| 1491 | /* Can only happen with delayed DSACK + discard craziness */ |
| 1492 | if (unlikely(skb == tcp_write_queue_head(sk))) |
| 1493 | goto fallback; |
| 1494 | prev = tcp_write_queue_prev(sk, skb); |
| 1495 | |
| 1496 | if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) |
| 1497 | goto fallback; |
| 1498 | |
| 1499 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && |
| 1500 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); |
| 1501 | |
| 1502 | if (in_sack) { |
| 1503 | len = skb->len; |
| 1504 | pcount = tcp_skb_pcount(skb); |
Ilpo Järvinen | 775ffab | 2008-12-05 22:41:26 -0800 | [diff] [blame] | 1505 | mss = tcp_skb_seglen(skb); |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1506 | |
| 1507 | /* TODO: Fix DSACKs to not fragment already SACKed and we can |
| 1508 | * drop this restriction as unnecessary |
| 1509 | */ |
Ilpo Järvinen | 775ffab | 2008-12-05 22:41:26 -0800 | [diff] [blame] | 1510 | if (mss != tcp_skb_seglen(prev)) |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1511 | goto fallback; |
| 1512 | } else { |
| 1513 | if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) |
| 1514 | goto noop; |
| 1515 | /* CHECKME: This is non-MSS split case only?, this will |
| 1516 | * cause skipped skbs due to advancing loop btw, original |
| 1517 | * has that feature too |
| 1518 | */ |
| 1519 | if (tcp_skb_pcount(skb) <= 1) |
| 1520 | goto noop; |
| 1521 | |
| 1522 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); |
| 1523 | if (!in_sack) { |
| 1524 | /* TODO: head merge to next could be attempted here |
| 1525 | * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), |
| 1526 | * though it might not be worth of the additional hassle |
| 1527 | * |
| 1528 | * ...we can probably just fallback to what was done |
| 1529 | * previously. We could try merging non-SACKed ones |
| 1530 | * as well but it probably isn't going to buy off |
| 1531 | * because later SACKs might again split them, and |
| 1532 | * it would make skb timestamp tracking considerably |
| 1533 | * harder problem. |
| 1534 | */ |
| 1535 | goto fallback; |
| 1536 | } |
| 1537 | |
| 1538 | len = end_seq - TCP_SKB_CB(skb)->seq; |
| 1539 | BUG_ON(len < 0); |
| 1540 | BUG_ON(len > skb->len); |
| 1541 | |
| 1542 | /* MSS boundaries should be honoured or else pcount will |
| 1543 | * severely break even though it makes things bit trickier. |
| 1544 | * Optimize common case to avoid most of the divides |
| 1545 | */ |
| 1546 | mss = tcp_skb_mss(skb); |
| 1547 | |
| 1548 | /* TODO: Fix DSACKs to not fragment already SACKed and we can |
| 1549 | * drop this restriction as unnecessary |
| 1550 | */ |
Ilpo Järvinen | 775ffab | 2008-12-05 22:41:26 -0800 | [diff] [blame] | 1551 | if (mss != tcp_skb_seglen(prev)) |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1552 | goto fallback; |
| 1553 | |
| 1554 | if (len == mss) { |
| 1555 | pcount = 1; |
| 1556 | } else if (len < mss) { |
| 1557 | goto noop; |
| 1558 | } else { |
| 1559 | pcount = len / mss; |
| 1560 | len = pcount * mss; |
| 1561 | } |
| 1562 | } |
| 1563 | |
| 1564 | if (!skb_shift(prev, skb, len)) |
| 1565 | goto fallback; |
Ilpo Järvinen | 9ec06ff | 2009-03-01 00:21:36 -0800 | [diff] [blame] | 1566 | if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack)) |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1567 | goto out; |
| 1568 | |
| 1569 | /* Hole filled allows collapsing with the next as well, this is very |
| 1570 | * useful when hole on every nth skb pattern happens |
| 1571 | */ |
| 1572 | if (prev == tcp_write_queue_tail(sk)) |
| 1573 | goto out; |
| 1574 | skb = tcp_write_queue_next(sk, prev); |
| 1575 | |
Ilpo Järvinen | f0bc52f | 2008-12-05 22:40:47 -0800 | [diff] [blame] | 1576 | if (!skb_can_shift(skb) || |
| 1577 | (skb == tcp_send_head(sk)) || |
| 1578 | ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || |
Ilpo Järvinen | 775ffab | 2008-12-05 22:41:26 -0800 | [diff] [blame] | 1579 | (mss != tcp_skb_seglen(skb))) |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1580 | goto out; |
| 1581 | |
| 1582 | len = skb->len; |
| 1583 | if (skb_shift(prev, skb, len)) { |
| 1584 | pcount += tcp_skb_pcount(skb); |
Ilpo Järvinen | 9ec06ff | 2009-03-01 00:21:36 -0800 | [diff] [blame] | 1585 | tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0); |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1586 | } |
| 1587 | |
| 1588 | out: |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1589 | state->fack_count += pcount; |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1590 | return prev; |
| 1591 | |
| 1592 | noop: |
| 1593 | return skb; |
| 1594 | |
| 1595 | fallback: |
Ilpo Järvinen | 111cc8b | 2008-11-24 21:27:22 -0800 | [diff] [blame] | 1596 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1597 | return NULL; |
| 1598 | } |
| 1599 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1600 | static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, |
| 1601 | struct tcp_sack_block *next_dup, |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1602 | struct tcp_sacktag_state *state, |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1603 | u32 start_seq, u32 end_seq, |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1604 | int dup_sack_in) |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1605 | { |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1606 | struct tcp_sock *tp = tcp_sk(sk); |
| 1607 | struct sk_buff *tmp; |
| 1608 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1609 | tcp_for_write_queue_from(skb, sk) { |
| 1610 | int in_sack = 0; |
| 1611 | int dup_sack = dup_sack_in; |
| 1612 | |
| 1613 | if (skb == tcp_send_head(sk)) |
| 1614 | break; |
| 1615 | |
| 1616 | /* queue is in-order => we can short-circuit the walk early */ |
| 1617 | if (!before(TCP_SKB_CB(skb)->seq, end_seq)) |
| 1618 | break; |
| 1619 | |
| 1620 | if ((next_dup != NULL) && |
| 1621 | before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { |
| 1622 | in_sack = tcp_match_skb_to_sack(sk, skb, |
| 1623 | next_dup->start_seq, |
| 1624 | next_dup->end_seq); |
| 1625 | if (in_sack > 0) |
| 1626 | dup_sack = 1; |
| 1627 | } |
| 1628 | |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1629 | /* skb reference here is a bit tricky to get right, since |
| 1630 | * shifting can eat and free both this skb and the next, |
| 1631 | * so not even _safe variant of the loop is enough. |
| 1632 | */ |
| 1633 | if (in_sack <= 0) { |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1634 | tmp = tcp_shift_skb_data(sk, skb, state, |
| 1635 | start_seq, end_seq, dup_sack); |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1636 | if (tmp != NULL) { |
| 1637 | if (tmp != skb) { |
| 1638 | skb = tmp; |
| 1639 | continue; |
| 1640 | } |
| 1641 | |
| 1642 | in_sack = 0; |
| 1643 | } else { |
| 1644 | in_sack = tcp_match_skb_to_sack(sk, skb, |
| 1645 | start_seq, |
| 1646 | end_seq); |
| 1647 | } |
| 1648 | } |
| 1649 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1650 | if (unlikely(in_sack < 0)) |
| 1651 | break; |
| 1652 | |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1653 | if (in_sack) { |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1654 | TCP_SKB_CB(skb)->sacked = tcp_sacktag_one(skb, sk, |
| 1655 | state, |
| 1656 | dup_sack, |
| 1657 | tcp_skb_pcount(skb)); |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1658 | |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1659 | if (!before(TCP_SKB_CB(skb)->seq, |
| 1660 | tcp_highest_sack_seq(tp))) |
| 1661 | tcp_advance_highest_sack(sk, skb); |
| 1662 | } |
| 1663 | |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1664 | state->fack_count += tcp_skb_pcount(skb); |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1665 | } |
| 1666 | return skb; |
| 1667 | } |
| 1668 | |
| 1669 | /* Avoid all extra work that is being done by sacktag while walking in |
| 1670 | * a normal way |
| 1671 | */ |
| 1672 | static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1673 | struct tcp_sacktag_state *state, |
| 1674 | u32 skip_to_seq) |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1675 | { |
| 1676 | tcp_for_write_queue_from(skb, sk) { |
| 1677 | if (skb == tcp_send_head(sk)) |
| 1678 | break; |
| 1679 | |
Ilpo Järvinen | e8bae27 | 2008-11-24 21:12:28 -0800 | [diff] [blame] | 1680 | if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq)) |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1681 | break; |
Ilpo Järvinen | d152a7d | 2008-03-03 12:10:16 -0800 | [diff] [blame] | 1682 | |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1683 | state->fack_count += tcp_skb_pcount(skb); |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1684 | } |
| 1685 | return skb; |
| 1686 | } |
| 1687 | |
| 1688 | static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, |
| 1689 | struct sock *sk, |
| 1690 | struct tcp_sack_block *next_dup, |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1691 | struct tcp_sacktag_state *state, |
| 1692 | u32 skip_to_seq) |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1693 | { |
| 1694 | if (next_dup == NULL) |
| 1695 | return skb; |
| 1696 | |
| 1697 | if (before(next_dup->start_seq, skip_to_seq)) { |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1698 | skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq); |
| 1699 | skb = tcp_sacktag_walk(skb, sk, NULL, state, |
| 1700 | next_dup->start_seq, next_dup->end_seq, |
| 1701 | 1); |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1702 | } |
| 1703 | |
| 1704 | return skb; |
| 1705 | } |
| 1706 | |
| 1707 | static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache) |
| 1708 | { |
| 1709 | return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); |
| 1710 | } |
| 1711 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1712 | static int |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1713 | tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, |
| 1714 | u32 prior_snd_una) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1715 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1716 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1717 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | 9c70220 | 2007-04-25 18:04:18 -0700 | [diff] [blame] | 1718 | unsigned char *ptr = (skb_transport_header(ack_skb) + |
| 1719 | TCP_SKB_CB(ack_skb)->sacked); |
Ilpo Järvinen | fd6dad6 | 2007-11-15 19:49:47 -0800 | [diff] [blame] | 1720 | struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); |
Adam Langley | 4389dde | 2008-07-19 00:07:02 -0700 | [diff] [blame] | 1721 | struct tcp_sack_block sp[TCP_NUM_SACKS]; |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1722 | struct tcp_sack_block *cache; |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1723 | struct tcp_sacktag_state state; |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1724 | struct sk_buff *skb; |
Adam Langley | 4389dde | 2008-07-19 00:07:02 -0700 | [diff] [blame] | 1725 | int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); |
Ilpo Järvinen | fd6dad6 | 2007-11-15 19:49:47 -0800 | [diff] [blame] | 1726 | int used_sacks; |
Ilpo Järvinen | 7769f40 | 2007-06-15 15:14:04 -0700 | [diff] [blame] | 1727 | int found_dup_sack = 0; |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1728 | int i, j; |
Baruch Even | fda03fb | 2007-02-04 23:35:57 -0800 | [diff] [blame] | 1729 | int first_sack_index; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1730 | |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1731 | state.flag = 0; |
| 1732 | state.reord = tp->packets_out; |
| 1733 | |
Ilpo Järvinen | d738cd8 | 2007-03-24 21:03:23 -0700 | [diff] [blame] | 1734 | if (!tp->sacked_out) { |
Ilpo Järvinen | de83c05 | 2007-10-07 23:37:55 -0700 | [diff] [blame] | 1735 | if (WARN_ON(tp->fackets_out)) |
| 1736 | tp->fackets_out = 0; |
Ilpo Järvinen | 6859d49 | 2007-12-02 00:48:06 +0200 | [diff] [blame] | 1737 | tcp_highest_sack_reset(sk); |
Ilpo Järvinen | d738cd8 | 2007-03-24 21:03:23 -0700 | [diff] [blame] | 1738 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1739 | |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 1740 | found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, |
David S. Miller | d06e021 | 2007-06-18 22:43:06 -0700 | [diff] [blame] | 1741 | num_sacks, prior_snd_una); |
| 1742 | if (found_dup_sack) |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1743 | state.flag |= FLAG_DSACKING_ACK; |
Baruch Even | 6f74651 | 2007-02-04 23:36:42 -0800 | [diff] [blame] | 1744 | |
| 1745 | /* Eliminate too old ACKs, but take into |
| 1746 | * account more or less fresh ones, they can |
| 1747 | * contain valid SACK info. |
| 1748 | */ |
| 1749 | if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) |
| 1750 | return 0; |
| 1751 | |
Ilpo Järvinen | 96a2d41 | 2007-11-14 15:47:18 -0800 | [diff] [blame] | 1752 | if (!tp->packets_out) |
| 1753 | goto out; |
| 1754 | |
Ilpo Järvinen | fd6dad6 | 2007-11-15 19:49:47 -0800 | [diff] [blame] | 1755 | used_sacks = 0; |
| 1756 | first_sack_index = 0; |
| 1757 | for (i = 0; i < num_sacks; i++) { |
| 1758 | int dup_sack = !i && found_dup_sack; |
| 1759 | |
Harvey Harrison | d3e2ce3 | 2008-05-02 16:26:16 -0700 | [diff] [blame] | 1760 | sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq); |
| 1761 | sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq); |
Ilpo Järvinen | fd6dad6 | 2007-11-15 19:49:47 -0800 | [diff] [blame] | 1762 | |
| 1763 | if (!tcp_is_sackblock_valid(tp, dup_sack, |
| 1764 | sp[used_sacks].start_seq, |
| 1765 | sp[used_sacks].end_seq)) { |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 1766 | int mib_idx; |
| 1767 | |
Ilpo Järvinen | fd6dad6 | 2007-11-15 19:49:47 -0800 | [diff] [blame] | 1768 | if (dup_sack) { |
| 1769 | if (!tp->undo_marker) |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 1770 | mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO; |
Ilpo Järvinen | fd6dad6 | 2007-11-15 19:49:47 -0800 | [diff] [blame] | 1771 | else |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 1772 | mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD; |
Ilpo Järvinen | fd6dad6 | 2007-11-15 19:49:47 -0800 | [diff] [blame] | 1773 | } else { |
| 1774 | /* Don't count olds caused by ACK reordering */ |
| 1775 | if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && |
| 1776 | !after(sp[used_sacks].end_seq, tp->snd_una)) |
| 1777 | continue; |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 1778 | mib_idx = LINUX_MIB_TCPSACKDISCARD; |
Ilpo Järvinen | fd6dad6 | 2007-11-15 19:49:47 -0800 | [diff] [blame] | 1779 | } |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 1780 | |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 1781 | NET_INC_STATS_BH(sock_net(sk), mib_idx); |
Ilpo Järvinen | fd6dad6 | 2007-11-15 19:49:47 -0800 | [diff] [blame] | 1782 | if (i == 0) |
| 1783 | first_sack_index = -1; |
| 1784 | continue; |
| 1785 | } |
| 1786 | |
| 1787 | /* Ignore very old stuff early */ |
| 1788 | if (!after(sp[used_sacks].end_seq, prior_snd_una)) |
| 1789 | continue; |
| 1790 | |
| 1791 | used_sacks++; |
| 1792 | } |
| 1793 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1794 | /* order SACK blocks to allow in order walk of the retrans queue */ |
| 1795 | for (i = used_sacks - 1; i > 0; i--) { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1796 | for (j = 0; j < i; j++) { |
| 1797 | if (after(sp[j].start_seq, sp[j + 1].start_seq)) { |
Ilpo Järvinen | a0bffff | 2009-03-21 13:36:17 -0700 | [diff] [blame] | 1798 | swap(sp[j], sp[j + 1]); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1799 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1800 | /* Track where the first SACK block goes to */ |
| 1801 | if (j == first_sack_index) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1802 | first_sack_index = j + 1; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1803 | } |
| 1804 | } |
| 1805 | } |
| 1806 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1807 | skb = tcp_write_queue_head(sk); |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1808 | state.fack_count = 0; |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1809 | i = 0; |
| 1810 | |
| 1811 | if (!tp->sacked_out) { |
| 1812 | /* It's already past, so skip checking against it */ |
| 1813 | cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); |
| 1814 | } else { |
| 1815 | cache = tp->recv_sack_cache; |
| 1816 | /* Skip empty blocks in at head of the cache */ |
| 1817 | while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && |
| 1818 | !cache->end_seq) |
| 1819 | cache++; |
Baruch Even | fda03fb | 2007-02-04 23:35:57 -0800 | [diff] [blame] | 1820 | } |
| 1821 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1822 | while (i < used_sacks) { |
Ilpo Järvinen | fd6dad6 | 2007-11-15 19:49:47 -0800 | [diff] [blame] | 1823 | u32 start_seq = sp[i].start_seq; |
| 1824 | u32 end_seq = sp[i].end_seq; |
Ilpo Järvinen | 7769f40 | 2007-06-15 15:14:04 -0700 | [diff] [blame] | 1825 | int dup_sack = (found_dup_sack && (i == first_sack_index)); |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1826 | struct tcp_sack_block *next_dup = NULL; |
Ilpo Järvinen | e56d6cd | 2007-11-01 00:09:37 -0700 | [diff] [blame] | 1827 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1828 | if (found_dup_sack && ((i + 1) == first_sack_index)) |
| 1829 | next_dup = &sp[i + 1]; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1830 | |
| 1831 | /* Event "B" in the comment above. */ |
| 1832 | if (after(end_seq, tp->high_seq)) |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1833 | state.flag |= FLAG_DATA_LOST; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1834 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1835 | /* Skip too early cached blocks */ |
| 1836 | while (tcp_sack_cache_ok(tp, cache) && |
| 1837 | !before(start_seq, cache->end_seq)) |
| 1838 | cache++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1839 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1840 | /* Can skip some work by looking recv_sack_cache? */ |
| 1841 | if (tcp_sack_cache_ok(tp, cache) && !dup_sack && |
| 1842 | after(end_seq, cache->start_seq)) { |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1843 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1844 | /* Head todo? */ |
| 1845 | if (before(start_seq, cache->start_seq)) { |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1846 | skb = tcp_sacktag_skip(skb, sk, &state, |
| 1847 | start_seq); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1848 | skb = tcp_sacktag_walk(skb, sk, next_dup, |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1849 | &state, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1850 | start_seq, |
| 1851 | cache->start_seq, |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1852 | dup_sack); |
Baruch Even | fda03fb | 2007-02-04 23:35:57 -0800 | [diff] [blame] | 1853 | } |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1854 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1855 | /* Rest of the block already fully processed? */ |
Ilpo Järvinen | 20de20b | 2007-11-16 16:17:05 -0800 | [diff] [blame] | 1856 | if (!after(end_seq, cache->end_seq)) |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1857 | goto advance_sp; |
Ilpo Järvinen | 20de20b | 2007-11-16 16:17:05 -0800 | [diff] [blame] | 1858 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1859 | skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1860 | &state, |
| 1861 | cache->end_seq); |
Ilpo Järvinen | e56d6cd | 2007-11-01 00:09:37 -0700 | [diff] [blame] | 1862 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1863 | /* ...tail remains todo... */ |
Ilpo Järvinen | 6859d49 | 2007-12-02 00:48:06 +0200 | [diff] [blame] | 1864 | if (tcp_highest_sack_seq(tp) == cache->end_seq) { |
Ilpo Järvinen | 20de20b | 2007-11-16 16:17:05 -0800 | [diff] [blame] | 1865 | /* ...but better entrypoint exists! */ |
Ilpo Järvinen | 6859d49 | 2007-12-02 00:48:06 +0200 | [diff] [blame] | 1866 | skb = tcp_highest_sack(sk); |
| 1867 | if (skb == NULL) |
| 1868 | break; |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1869 | state.fack_count = tp->fackets_out; |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1870 | cache++; |
| 1871 | goto walk; |
| 1872 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1873 | |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1874 | skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq); |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1875 | /* Check overlap against next cached too (past this one already) */ |
| 1876 | cache++; |
| 1877 | continue; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1878 | } |
Ilpo Järvinen | fbd52eb | 2007-11-10 21:24:19 -0800 | [diff] [blame] | 1879 | |
Ilpo Järvinen | 6859d49 | 2007-12-02 00:48:06 +0200 | [diff] [blame] | 1880 | if (!before(start_seq, tcp_highest_sack_seq(tp))) { |
| 1881 | skb = tcp_highest_sack(sk); |
| 1882 | if (skb == NULL) |
| 1883 | break; |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1884 | state.fack_count = tp->fackets_out; |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1885 | } |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1886 | skb = tcp_sacktag_skip(skb, sk, &state, start_seq); |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1887 | |
| 1888 | walk: |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1889 | skb = tcp_sacktag_walk(skb, sk, next_dup, &state, |
| 1890 | start_seq, end_seq, dup_sack); |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1891 | |
| 1892 | advance_sp: |
Ilpo Järvinen | fbd52eb | 2007-11-10 21:24:19 -0800 | [diff] [blame] | 1893 | /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct |
| 1894 | * due to in-order walk |
| 1895 | */ |
| 1896 | if (after(end_seq, tp->frto_highmark)) |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1897 | state.flag &= ~FLAG_ONLY_ORIG_SACKED; |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1898 | |
| 1899 | i++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1900 | } |
| 1901 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1902 | /* Clear the head of the cache sack blocks so we can skip it next time */ |
| 1903 | for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { |
| 1904 | tp->recv_sack_cache[i].start_seq = 0; |
| 1905 | tp->recv_sack_cache[i].end_seq = 0; |
| 1906 | } |
| 1907 | for (j = 0; j < used_sacks; j++) |
| 1908 | tp->recv_sack_cache[i++] = sp[j]; |
| 1909 | |
Ilpo Järvinen | 407ef1d | 2007-12-02 00:47:57 +0200 | [diff] [blame] | 1910 | tcp_mark_lost_retrans(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1911 | |
Ilpo Järvinen | 86426c2 | 2007-08-09 14:45:17 +0300 | [diff] [blame] | 1912 | tcp_verify_left_out(tp); |
| 1913 | |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1914 | if ((state.reord < tp->fackets_out) && |
Ilpo Järvinen | f577111 | 2007-11-15 19:35:11 -0800 | [diff] [blame] | 1915 | ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) && |
Ilpo Järvinen | c5e7af0 | 2007-02-23 16:22:06 -0800 | [diff] [blame] | 1916 | (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark))) |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1917 | tcp_update_reordering(sk, tp->fackets_out - state.reord, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1918 | |
Ilpo Järvinen | 96a2d41 | 2007-11-14 15:47:18 -0800 | [diff] [blame] | 1919 | out: |
| 1920 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1921 | #if FASTRETRANS_DEBUG > 0 |
Ilpo Järvinen | 547b792 | 2008-07-25 21:43:18 -0700 | [diff] [blame] | 1922 | WARN_ON((int)tp->sacked_out < 0); |
| 1923 | WARN_ON((int)tp->lost_out < 0); |
| 1924 | WARN_ON((int)tp->retrans_out < 0); |
| 1925 | WARN_ON((int)tcp_packets_in_flight(tp) < 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1926 | #endif |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1927 | return state.flag; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1928 | } |
| 1929 | |
Ilpo Järvinen | 882beba | 2008-04-07 22:33:07 -0700 | [diff] [blame] | 1930 | /* Limits sacked_out so that sum with lost_out isn't ever larger than |
| 1931 | * packets_out. Returns zero if sacked_out adjustement wasn't necessary. |
Ilpo Järvinen | 30935cf | 2007-02-21 23:01:36 -0800 | [diff] [blame] | 1932 | */ |
Ilpo Järvinen | 8eecaba | 2008-11-25 13:45:29 -0800 | [diff] [blame] | 1933 | static int tcp_limit_reno_sacked(struct tcp_sock *tp) |
Ilpo Järvinen | 4ddf667 | 2007-05-27 01:52:00 -0700 | [diff] [blame] | 1934 | { |
Ilpo Järvinen | 4ddf667 | 2007-05-27 01:52:00 -0700 | [diff] [blame] | 1935 | u32 holes; |
| 1936 | |
| 1937 | holes = max(tp->lost_out, 1U); |
| 1938 | holes = min(holes, tp->packets_out); |
| 1939 | |
| 1940 | if ((tp->sacked_out + holes) > tp->packets_out) { |
| 1941 | tp->sacked_out = tp->packets_out - holes; |
Ilpo Järvinen | 882beba | 2008-04-07 22:33:07 -0700 | [diff] [blame] | 1942 | return 1; |
Ilpo Järvinen | 4ddf667 | 2007-05-27 01:52:00 -0700 | [diff] [blame] | 1943 | } |
Ilpo Järvinen | 882beba | 2008-04-07 22:33:07 -0700 | [diff] [blame] | 1944 | return 0; |
| 1945 | } |
| 1946 | |
| 1947 | /* If we receive more dupacks than we expected counting segments |
| 1948 | * in assumption of absent reordering, interpret this as reordering. |
| 1949 | * The only another reason could be bug in receiver TCP. |
| 1950 | */ |
| 1951 | static void tcp_check_reno_reordering(struct sock *sk, const int addend) |
| 1952 | { |
| 1953 | struct tcp_sock *tp = tcp_sk(sk); |
| 1954 | if (tcp_limit_reno_sacked(tp)) |
| 1955 | tcp_update_reordering(sk, tp->packets_out + addend, 0); |
Ilpo Järvinen | 4ddf667 | 2007-05-27 01:52:00 -0700 | [diff] [blame] | 1956 | } |
| 1957 | |
| 1958 | /* Emulate SACKs for SACKless connection: account for a new dupack. */ |
| 1959 | |
| 1960 | static void tcp_add_reno_sack(struct sock *sk) |
| 1961 | { |
| 1962 | struct tcp_sock *tp = tcp_sk(sk); |
| 1963 | tp->sacked_out++; |
| 1964 | tcp_check_reno_reordering(sk, 0); |
Ilpo Järvinen | 005903b | 2007-08-09 14:44:16 +0300 | [diff] [blame] | 1965 | tcp_verify_left_out(tp); |
Ilpo Järvinen | 4ddf667 | 2007-05-27 01:52:00 -0700 | [diff] [blame] | 1966 | } |
| 1967 | |
| 1968 | /* Account for ACK, ACKing some data in Reno Recovery phase. */ |
| 1969 | |
| 1970 | static void tcp_remove_reno_sacks(struct sock *sk, int acked) |
| 1971 | { |
| 1972 | struct tcp_sock *tp = tcp_sk(sk); |
| 1973 | |
| 1974 | if (acked > 0) { |
| 1975 | /* One ACK acked hole. The rest eat duplicate ACKs. */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1976 | if (acked - 1 >= tp->sacked_out) |
Ilpo Järvinen | 4ddf667 | 2007-05-27 01:52:00 -0700 | [diff] [blame] | 1977 | tp->sacked_out = 0; |
| 1978 | else |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1979 | tp->sacked_out -= acked - 1; |
Ilpo Järvinen | 4ddf667 | 2007-05-27 01:52:00 -0700 | [diff] [blame] | 1980 | } |
| 1981 | tcp_check_reno_reordering(sk, acked); |
Ilpo Järvinen | 005903b | 2007-08-09 14:44:16 +0300 | [diff] [blame] | 1982 | tcp_verify_left_out(tp); |
Ilpo Järvinen | 4ddf667 | 2007-05-27 01:52:00 -0700 | [diff] [blame] | 1983 | } |
| 1984 | |
| 1985 | static inline void tcp_reset_reno_sack(struct tcp_sock *tp) |
| 1986 | { |
| 1987 | tp->sacked_out = 0; |
Ilpo Järvinen | 4ddf667 | 2007-05-27 01:52:00 -0700 | [diff] [blame] | 1988 | } |
| 1989 | |
Ilpo Järvinen | 62ab222 | 2008-05-08 01:09:11 -0700 | [diff] [blame] | 1990 | static int tcp_is_sackfrto(const struct tcp_sock *tp) |
| 1991 | { |
| 1992 | return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp); |
| 1993 | } |
| 1994 | |
Ilpo Järvinen | 95eacd2 | 2007-10-01 15:27:42 -0700 | [diff] [blame] | 1995 | /* F-RTO can only be used if TCP has never retransmitted anything other than |
| 1996 | * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here) |
| 1997 | */ |
Ilpo Järvinen | 46d0de4 | 2007-02-21 23:10:39 -0800 | [diff] [blame] | 1998 | int tcp_use_frto(struct sock *sk) |
Ilpo Järvinen | bdaae17 | 2007-02-21 22:59:58 -0800 | [diff] [blame] | 1999 | { |
| 2000 | const struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 6adb4f7 | 2008-04-07 22:33:57 -0700 | [diff] [blame] | 2001 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Ilpo Järvinen | 46d0de4 | 2007-02-21 23:10:39 -0800 | [diff] [blame] | 2002 | struct sk_buff *skb; |
Ilpo Järvinen | bdaae17 | 2007-02-21 22:59:58 -0800 | [diff] [blame] | 2003 | |
Ilpo Järvinen | 575ee71 | 2007-04-30 00:39:55 -0700 | [diff] [blame] | 2004 | if (!sysctl_tcp_frto) |
Ilpo Järvinen | 46d0de4 | 2007-02-21 23:10:39 -0800 | [diff] [blame] | 2005 | return 0; |
| 2006 | |
Ilpo Järvinen | 6adb4f7 | 2008-04-07 22:33:57 -0700 | [diff] [blame] | 2007 | /* MTU probe and F-RTO won't really play nicely along currently */ |
| 2008 | if (icsk->icsk_mtup.probe_size) |
| 2009 | return 0; |
| 2010 | |
Ilpo Järvinen | 62ab222 | 2008-05-08 01:09:11 -0700 | [diff] [blame] | 2011 | if (tcp_is_sackfrto(tp)) |
Ilpo Järvinen | 4dc2665 | 2007-02-21 23:16:11 -0800 | [diff] [blame] | 2012 | return 1; |
| 2013 | |
Ilpo Järvinen | 46d0de4 | 2007-02-21 23:10:39 -0800 | [diff] [blame] | 2014 | /* Avoid expensive walking of rexmit queue if possible */ |
| 2015 | if (tp->retrans_out > 1) |
| 2016 | return 0; |
| 2017 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2018 | skb = tcp_write_queue_head(sk); |
David S. Miller | 28e3487 | 2008-09-23 02:51:41 -0700 | [diff] [blame] | 2019 | if (tcp_skb_is_last(sk, skb)) |
| 2020 | return 1; |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2021 | skb = tcp_write_queue_next(sk, skb); /* Skips head */ |
| 2022 | tcp_for_write_queue_from(skb, sk) { |
| 2023 | if (skb == tcp_send_head(sk)) |
| 2024 | break; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2025 | if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) |
Ilpo Järvinen | 46d0de4 | 2007-02-21 23:10:39 -0800 | [diff] [blame] | 2026 | return 0; |
| 2027 | /* Short-circuit when first non-SACKed skb has been checked */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2028 | if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) |
Ilpo Järvinen | 46d0de4 | 2007-02-21 23:10:39 -0800 | [diff] [blame] | 2029 | break; |
| 2030 | } |
| 2031 | return 1; |
Ilpo Järvinen | bdaae17 | 2007-02-21 22:59:58 -0800 | [diff] [blame] | 2032 | } |
| 2033 | |
Ilpo Järvinen | 30935cf | 2007-02-21 23:01:36 -0800 | [diff] [blame] | 2034 | /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO |
| 2035 | * recovery a bit and use heuristics in tcp_process_frto() to detect if |
Ilpo Järvinen | d1a54c6 | 2007-02-21 23:11:57 -0800 | [diff] [blame] | 2036 | * the RTO was spurious. Only clear SACKED_RETRANS of the head here to |
| 2037 | * keep retrans_out counting accurate (with SACK F-RTO, other than head |
| 2038 | * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS |
| 2039 | * bits are handled if the Loss state is really to be entered (in |
| 2040 | * tcp_enter_frto_loss). |
Ilpo Järvinen | 7487c48 | 2007-02-21 23:02:30 -0800 | [diff] [blame] | 2041 | * |
| 2042 | * Do like tcp_enter_loss() would; when RTO expires the second time it |
| 2043 | * does: |
| 2044 | * "Reduce ssthresh if it has not yet been made inside this window." |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2045 | */ |
| 2046 | void tcp_enter_frto(struct sock *sk) |
| 2047 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2048 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2049 | struct tcp_sock *tp = tcp_sk(sk); |
| 2050 | struct sk_buff *skb; |
| 2051 | |
Ilpo Järvinen | 7487c48 | 2007-02-21 23:02:30 -0800 | [diff] [blame] | 2052 | if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) || |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2053 | tp->snd_una == tp->high_seq || |
Ilpo Järvinen | 7487c48 | 2007-02-21 23:02:30 -0800 | [diff] [blame] | 2054 | ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) && |
| 2055 | !icsk->icsk_retransmits)) { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2056 | tp->prior_ssthresh = tcp_current_ssthresh(sk); |
Ilpo Järvinen | 66e93e4 | 2007-02-21 23:13:47 -0800 | [diff] [blame] | 2057 | /* Our state is too optimistic in ssthresh() call because cwnd |
Ryousei Takano | 564262c1 | 2007-10-25 23:03:52 -0700 | [diff] [blame] | 2058 | * is not reduced until tcp_enter_frto_loss() when previous F-RTO |
Ilpo Järvinen | 66e93e4 | 2007-02-21 23:13:47 -0800 | [diff] [blame] | 2059 | * recovery has not yet completed. Pattern would be this: RTO, |
| 2060 | * Cumulative ACK, RTO (2xRTO for the same segment does not end |
| 2061 | * up here twice). |
| 2062 | * RFC4138 should be more specific on what to do, even though |
| 2063 | * RTO is quite unlikely to occur after the first Cumulative ACK |
| 2064 | * due to back-off and complexity of triggering events ... |
| 2065 | */ |
| 2066 | if (tp->frto_counter) { |
| 2067 | u32 stored_cwnd; |
| 2068 | stored_cwnd = tp->snd_cwnd; |
| 2069 | tp->snd_cwnd = 2; |
| 2070 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); |
| 2071 | tp->snd_cwnd = stored_cwnd; |
| 2072 | } else { |
| 2073 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); |
| 2074 | } |
| 2075 | /* ... in theory, cong.control module could do "any tricks" in |
| 2076 | * ssthresh(), which means that ca_state, lost bits and lost_out |
| 2077 | * counter would have to be faked before the call occurs. We |
| 2078 | * consider that too expensive, unlikely and hacky, so modules |
| 2079 | * using these in ssthresh() must deal these incompatibility |
| 2080 | * issues if they receives CA_EVENT_FRTO and frto_counter != 0 |
| 2081 | */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2082 | tcp_ca_event(sk, CA_EVENT_FRTO); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2083 | } |
| 2084 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2085 | tp->undo_marker = tp->snd_una; |
| 2086 | tp->undo_retrans = 0; |
| 2087 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2088 | skb = tcp_write_queue_head(sk); |
Ilpo Järvinen | 009a2e3 | 2007-09-20 11:34:38 -0700 | [diff] [blame] | 2089 | if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) |
| 2090 | tp->undo_marker = 0; |
Ilpo Järvinen | d1a54c6 | 2007-02-21 23:11:57 -0800 | [diff] [blame] | 2091 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
Ilpo Järvinen | 522e754 | 2007-02-21 22:54:52 -0800 | [diff] [blame] | 2092 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; |
Ilpo Järvinen | d1a54c6 | 2007-02-21 23:11:57 -0800 | [diff] [blame] | 2093 | tp->retrans_out -= tcp_skb_pcount(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2094 | } |
Ilpo Järvinen | 005903b | 2007-08-09 14:44:16 +0300 | [diff] [blame] | 2095 | tcp_verify_left_out(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2096 | |
Ilpo Järvinen | 746aa32 | 2007-11-13 21:01:23 -0800 | [diff] [blame] | 2097 | /* Too bad if TCP was application limited */ |
| 2098 | tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1); |
| 2099 | |
Ilpo Järvinen | 4dc2665 | 2007-02-21 23:16:11 -0800 | [diff] [blame] | 2100 | /* Earlier loss recovery underway (see RFC4138; Appendix B). |
| 2101 | * The last condition is necessary at least in tp->frto_counter case. |
| 2102 | */ |
Ilpo Järvinen | 62ab222 | 2008-05-08 01:09:11 -0700 | [diff] [blame] | 2103 | if (tcp_is_sackfrto(tp) && (tp->frto_counter || |
Ilpo Järvinen | 4dc2665 | 2007-02-21 23:16:11 -0800 | [diff] [blame] | 2104 | ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) && |
| 2105 | after(tp->high_seq, tp->snd_una)) { |
| 2106 | tp->frto_highmark = tp->high_seq; |
| 2107 | } else { |
| 2108 | tp->frto_highmark = tp->snd_nxt; |
| 2109 | } |
Ilpo Järvinen | 7b0eb22 | 2007-02-21 23:03:35 -0800 | [diff] [blame] | 2110 | tcp_set_ca_state(sk, TCP_CA_Disorder); |
| 2111 | tp->high_seq = tp->snd_nxt; |
Ilpo Järvinen | 7487c48 | 2007-02-21 23:02:30 -0800 | [diff] [blame] | 2112 | tp->frto_counter = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2113 | } |
| 2114 | |
| 2115 | /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, |
| 2116 | * which indicates that we should follow the traditional RTO recovery, |
| 2117 | * i.e. mark everything lost and do go-back-N retransmission. |
| 2118 | */ |
Ilpo Järvinen | d1a54c6 | 2007-02-21 23:11:57 -0800 | [diff] [blame] | 2119 | static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2120 | { |
| 2121 | struct tcp_sock *tp = tcp_sk(sk); |
| 2122 | struct sk_buff *skb; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2123 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2124 | tp->lost_out = 0; |
Ilpo Järvinen | d1a54c6 | 2007-02-21 23:11:57 -0800 | [diff] [blame] | 2125 | tp->retrans_out = 0; |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 2126 | if (tcp_is_reno(tp)) |
Ilpo Järvinen | 9bff40f | 2007-05-27 01:53:49 -0700 | [diff] [blame] | 2127 | tcp_reset_reno_sack(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2128 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2129 | tcp_for_write_queue(skb, sk) { |
| 2130 | if (skb == tcp_send_head(sk)) |
| 2131 | break; |
Ilpo Järvinen | 23aeeec | 2007-11-13 21:03:13 -0800 | [diff] [blame] | 2132 | |
| 2133 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; |
Ilpo Järvinen | d1a54c6 | 2007-02-21 23:11:57 -0800 | [diff] [blame] | 2134 | /* |
| 2135 | * Count the retransmission made on RTO correctly (only when |
| 2136 | * waiting for the first ACK and did not get it)... |
| 2137 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2138 | if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) { |
Ilpo Järvinen | 0a9f2a4 | 2007-07-15 00:19:29 -0700 | [diff] [blame] | 2139 | /* For some reason this R-bit might get cleared? */ |
| 2140 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) |
| 2141 | tp->retrans_out += tcp_skb_pcount(skb); |
Ilpo Järvinen | d1a54c6 | 2007-02-21 23:11:57 -0800 | [diff] [blame] | 2142 | /* ...enter this if branch just for the first segment */ |
| 2143 | flag |= FLAG_DATA_ACKED; |
| 2144 | } else { |
Ilpo Järvinen | 009a2e3 | 2007-09-20 11:34:38 -0700 | [diff] [blame] | 2145 | if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) |
| 2146 | tp->undo_marker = 0; |
Ilpo Järvinen | 23aeeec | 2007-11-13 21:03:13 -0800 | [diff] [blame] | 2147 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; |
Ilpo Järvinen | d1a54c6 | 2007-02-21 23:11:57 -0800 | [diff] [blame] | 2148 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2149 | |
Ilpo Järvinen | 79d44516 | 2008-05-13 02:54:19 -0700 | [diff] [blame] | 2150 | /* Marking forward transmissions that were made after RTO lost |
| 2151 | * can cause unnecessary retransmissions in some scenarios, |
| 2152 | * SACK blocks will mitigate that in some but not in all cases. |
| 2153 | * We used to not mark them but it was causing break-ups with |
| 2154 | * receivers that do only in-order receival. |
| 2155 | * |
| 2156 | * TODO: we could detect presence of such receiver and select |
| 2157 | * different behavior per flow. |
| 2158 | */ |
| 2159 | if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { |
Ilpo Järvinen | 9bff40f | 2007-05-27 01:53:49 -0700 | [diff] [blame] | 2160 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; |
| 2161 | tp->lost_out += tcp_skb_pcount(skb); |
Ilpo Järvinen | 006f582 | 2008-09-20 21:20:20 -0700 | [diff] [blame] | 2162 | tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2163 | } |
| 2164 | } |
Ilpo Järvinen | 005903b | 2007-08-09 14:44:16 +0300 | [diff] [blame] | 2165 | tcp_verify_left_out(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2166 | |
Ilpo Järvinen | 95c4922 | 2007-02-21 23:05:18 -0800 | [diff] [blame] | 2167 | tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2168 | tp->snd_cwnd_cnt = 0; |
| 2169 | tp->snd_cwnd_stamp = tcp_time_stamp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2170 | tp->frto_counter = 0; |
Ilpo Järvinen | 16e9068 | 2007-10-11 17:32:31 -0700 | [diff] [blame] | 2171 | tp->bytes_acked = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2172 | |
| 2173 | tp->reordering = min_t(unsigned int, tp->reordering, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2174 | sysctl_tcp_reordering); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2175 | tcp_set_ca_state(sk, TCP_CA_Loss); |
Ilpo Järvinen | 79d44516 | 2008-05-13 02:54:19 -0700 | [diff] [blame] | 2176 | tp->high_seq = tp->snd_nxt; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2177 | TCP_ECN_queue_cwr(tp); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2178 | |
Ilpo Järvinen | 64edc27 | 2008-09-20 21:18:32 -0700 | [diff] [blame] | 2179 | tcp_clear_all_retrans_hints(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2180 | } |
| 2181 | |
Ilpo Järvinen | 4cd8299 | 2007-10-11 17:34:57 -0700 | [diff] [blame] | 2182 | static void tcp_clear_retrans_partial(struct tcp_sock *tp) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2183 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2184 | tp->retrans_out = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2185 | tp->lost_out = 0; |
| 2186 | |
| 2187 | tp->undo_marker = 0; |
| 2188 | tp->undo_retrans = 0; |
| 2189 | } |
| 2190 | |
Ilpo Järvinen | 4cd8299 | 2007-10-11 17:34:57 -0700 | [diff] [blame] | 2191 | void tcp_clear_retrans(struct tcp_sock *tp) |
| 2192 | { |
| 2193 | tcp_clear_retrans_partial(tp); |
| 2194 | |
| 2195 | tp->fackets_out = 0; |
| 2196 | tp->sacked_out = 0; |
| 2197 | } |
| 2198 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2199 | /* Enter Loss state. If "how" is not zero, forget all SACK information |
| 2200 | * and reset tags completely, otherwise preserve SACKs. If receiver |
| 2201 | * dropped its ofo queue, we will know this due to reneging detection. |
| 2202 | */ |
| 2203 | void tcp_enter_loss(struct sock *sk, int how) |
| 2204 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2205 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2206 | struct tcp_sock *tp = tcp_sk(sk); |
| 2207 | struct sk_buff *skb; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2208 | |
| 2209 | /* Reduce ssthresh if it has not yet been made inside this window. */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2210 | if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq || |
| 2211 | (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { |
| 2212 | tp->prior_ssthresh = tcp_current_ssthresh(sk); |
| 2213 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); |
| 2214 | tcp_ca_event(sk, CA_EVENT_LOSS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2215 | } |
| 2216 | tp->snd_cwnd = 1; |
| 2217 | tp->snd_cwnd_cnt = 0; |
| 2218 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 2219 | |
Stephen Hemminger | 9772efb | 2005-11-10 17:09:53 -0800 | [diff] [blame] | 2220 | tp->bytes_acked = 0; |
Ilpo Järvinen | 4cd8299 | 2007-10-11 17:34:57 -0700 | [diff] [blame] | 2221 | tcp_clear_retrans_partial(tp); |
| 2222 | |
| 2223 | if (tcp_is_reno(tp)) |
| 2224 | tcp_reset_reno_sack(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2225 | |
Ilpo Järvinen | b768920 | 2007-09-20 11:37:19 -0700 | [diff] [blame] | 2226 | if (!how) { |
| 2227 | /* Push undo marker, if it was plain RTO and nothing |
| 2228 | * was retransmitted. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2229 | tp->undo_marker = tp->snd_una; |
Ilpo Järvinen | b768920 | 2007-09-20 11:37:19 -0700 | [diff] [blame] | 2230 | } else { |
Ilpo Järvinen | 4cd8299 | 2007-10-11 17:34:57 -0700 | [diff] [blame] | 2231 | tp->sacked_out = 0; |
| 2232 | tp->fackets_out = 0; |
Ilpo Järvinen | b768920 | 2007-09-20 11:37:19 -0700 | [diff] [blame] | 2233 | } |
Ilpo Järvinen | 64edc27 | 2008-09-20 21:18:32 -0700 | [diff] [blame] | 2234 | tcp_clear_all_retrans_hints(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2235 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2236 | tcp_for_write_queue(skb, sk) { |
| 2237 | if (skb == tcp_send_head(sk)) |
| 2238 | break; |
Ilpo Järvinen | 4cd8299 | 2007-10-11 17:34:57 -0700 | [diff] [blame] | 2239 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2240 | if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2241 | tp->undo_marker = 0; |
| 2242 | TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; |
| 2243 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) { |
| 2244 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; |
| 2245 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; |
| 2246 | tp->lost_out += tcp_skb_pcount(skb); |
Ilpo Järvinen | 006f582 | 2008-09-20 21:20:20 -0700 | [diff] [blame] | 2247 | tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2248 | } |
| 2249 | } |
Ilpo Järvinen | 005903b | 2007-08-09 14:44:16 +0300 | [diff] [blame] | 2250 | tcp_verify_left_out(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2251 | |
| 2252 | tp->reordering = min_t(unsigned int, tp->reordering, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2253 | sysctl_tcp_reordering); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2254 | tcp_set_ca_state(sk, TCP_CA_Loss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2255 | tp->high_seq = tp->snd_nxt; |
| 2256 | TCP_ECN_queue_cwr(tp); |
Ryousei Takano | 564262c1 | 2007-10-25 23:03:52 -0700 | [diff] [blame] | 2257 | /* Abort F-RTO algorithm if one is in progress */ |
Ilpo Järvinen | 580e572 | 2007-05-19 13:56:57 -0700 | [diff] [blame] | 2258 | tp->frto_counter = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2259 | } |
| 2260 | |
Ilpo Järvinen | cadbd03 | 2007-12-31 04:49:21 -0800 | [diff] [blame] | 2261 | /* If ACK arrived pointing to a remembered SACK, it means that our |
| 2262 | * remembered SACKs do not reflect real state of receiver i.e. |
| 2263 | * receiver _host_ is heavily congested (or buggy). |
| 2264 | * |
| 2265 | * Do processing similar to RTO timeout. |
| 2266 | */ |
| 2267 | static int tcp_check_sack_reneging(struct sock *sk, int flag) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2268 | { |
Ilpo Järvinen | cadbd03 | 2007-12-31 04:49:21 -0800 | [diff] [blame] | 2269 | if (flag & FLAG_SACK_RENEGING) { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2270 | struct inet_connection_sock *icsk = inet_csk(sk); |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 2271 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2272 | |
| 2273 | tcp_enter_loss(sk, 1); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2274 | icsk->icsk_retransmits++; |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2275 | tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2276 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2277 | icsk->icsk_rto, TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2278 | return 1; |
| 2279 | } |
| 2280 | return 0; |
| 2281 | } |
| 2282 | |
| 2283 | static inline int tcp_fackets_out(struct tcp_sock *tp) |
| 2284 | { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2285 | return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2286 | } |
| 2287 | |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2288 | /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs |
| 2289 | * counter when SACK is enabled (without SACK, sacked_out is used for |
| 2290 | * that purpose). |
| 2291 | * |
| 2292 | * Instead, with FACK TCP uses fackets_out that includes both SACKed |
| 2293 | * segments up to the highest received SACK block so far and holes in |
| 2294 | * between them. |
| 2295 | * |
| 2296 | * With reordering, holes may still be in flight, so RFC3517 recovery |
| 2297 | * uses pure sacked_out (total number of SACKed segments) even though |
| 2298 | * it violates the RFC that uses duplicate ACKs, often these are equal |
| 2299 | * but when e.g. out-of-window ACKs or packet duplication occurs, |
| 2300 | * they differ. Since neither occurs due to loss, TCP should really |
| 2301 | * ignore them. |
| 2302 | */ |
Andreas Petlund | ea84e55 | 2009-10-27 03:27:21 +0000 | [diff] [blame] | 2303 | static inline int tcp_dupack_heuristics(struct tcp_sock *tp) |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2304 | { |
| 2305 | return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1; |
| 2306 | } |
| 2307 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2308 | static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2309 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2310 | return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2311 | } |
| 2312 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2313 | static inline int tcp_head_timedout(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2314 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2315 | struct tcp_sock *tp = tcp_sk(sk); |
| 2316 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2317 | return tp->packets_out && |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2318 | tcp_skb_timedout(sk, tcp_write_queue_head(sk)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2319 | } |
| 2320 | |
| 2321 | /* Linux NewReno/SACK/FACK/ECN state machine. |
| 2322 | * -------------------------------------- |
| 2323 | * |
| 2324 | * "Open" Normal state, no dubious events, fast path. |
| 2325 | * "Disorder" In all the respects it is "Open", |
| 2326 | * but requires a bit more attention. It is entered when |
| 2327 | * we see some SACKs or dupacks. It is split of "Open" |
| 2328 | * mainly to move some processing from fast path to slow one. |
| 2329 | * "CWR" CWND was reduced due to some Congestion Notification event. |
| 2330 | * It can be ECN, ICMP source quench, local device congestion. |
| 2331 | * "Recovery" CWND was reduced, we are fast-retransmitting. |
| 2332 | * "Loss" CWND was reduced due to RTO timeout or SACK reneging. |
| 2333 | * |
| 2334 | * tcp_fastretrans_alert() is entered: |
| 2335 | * - each incoming ACK, if state is not "Open" |
| 2336 | * - when arrived ACK is unusual, namely: |
| 2337 | * * SACK |
| 2338 | * * Duplicate ACK. |
| 2339 | * * ECN ECE. |
| 2340 | * |
| 2341 | * Counting packets in flight is pretty simple. |
| 2342 | * |
| 2343 | * in_flight = packets_out - left_out + retrans_out |
| 2344 | * |
| 2345 | * packets_out is SND.NXT-SND.UNA counted in packets. |
| 2346 | * |
| 2347 | * retrans_out is number of retransmitted segments. |
| 2348 | * |
| 2349 | * left_out is number of segments left network, but not ACKed yet. |
| 2350 | * |
| 2351 | * left_out = sacked_out + lost_out |
| 2352 | * |
| 2353 | * sacked_out: Packets, which arrived to receiver out of order |
| 2354 | * and hence not ACKed. With SACKs this number is simply |
| 2355 | * amount of SACKed data. Even without SACKs |
| 2356 | * it is easy to give pretty reliable estimate of this number, |
| 2357 | * counting duplicate ACKs. |
| 2358 | * |
| 2359 | * lost_out: Packets lost by network. TCP has no explicit |
| 2360 | * "loss notification" feedback from network (for now). |
| 2361 | * It means that this number can be only _guessed_. |
| 2362 | * Actually, it is the heuristics to predict lossage that |
| 2363 | * distinguishes different algorithms. |
| 2364 | * |
| 2365 | * F.e. after RTO, when all the queue is considered as lost, |
| 2366 | * lost_out = packets_out and in_flight = retrans_out. |
| 2367 | * |
| 2368 | * Essentially, we have now two algorithms counting |
| 2369 | * lost packets. |
| 2370 | * |
| 2371 | * FACK: It is the simplest heuristics. As soon as we decided |
| 2372 | * that something is lost, we decide that _all_ not SACKed |
| 2373 | * packets until the most forward SACK are lost. I.e. |
| 2374 | * lost_out = fackets_out - sacked_out and left_out = fackets_out. |
| 2375 | * It is absolutely correct estimate, if network does not reorder |
| 2376 | * packets. And it loses any connection to reality when reordering |
| 2377 | * takes place. We use FACK by default until reordering |
| 2378 | * is suspected on the path to this destination. |
| 2379 | * |
| 2380 | * NewReno: when Recovery is entered, we assume that one segment |
| 2381 | * is lost (classic Reno). While we are in Recovery and |
| 2382 | * a partial ACK arrives, we assume that one more packet |
| 2383 | * is lost (NewReno). This heuristics are the same in NewReno |
| 2384 | * and SACK. |
| 2385 | * |
| 2386 | * Imagine, that's all! Forget about all this shamanism about CWND inflation |
| 2387 | * deflation etc. CWND is real congestion window, never inflated, changes |
| 2388 | * only according to classic VJ rules. |
| 2389 | * |
| 2390 | * Really tricky (and requiring careful tuning) part of algorithm |
| 2391 | * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). |
| 2392 | * The first determines the moment _when_ we should reduce CWND and, |
| 2393 | * hence, slow down forward transmission. In fact, it determines the moment |
| 2394 | * when we decide that hole is caused by loss, rather than by a reorder. |
| 2395 | * |
| 2396 | * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill |
| 2397 | * holes, caused by lost packets. |
| 2398 | * |
| 2399 | * And the most logically complicated part of algorithm is undo |
| 2400 | * heuristics. We detect false retransmits due to both too early |
| 2401 | * fast retransmit (reordering) and underestimated RTO, analyzing |
| 2402 | * timestamps and D-SACKs. When we detect that some segments were |
| 2403 | * retransmitted by mistake and CWND reduction was wrong, we undo |
| 2404 | * window reduction and abort recovery phase. This logic is hidden |
| 2405 | * inside several functions named tcp_try_undo_<something>. |
| 2406 | */ |
| 2407 | |
| 2408 | /* This function decides, when we should leave Disordered state |
| 2409 | * and enter Recovery phase, reducing congestion window. |
| 2410 | * |
| 2411 | * Main question: may we further continue forward transmission |
| 2412 | * with the same cwnd? |
| 2413 | */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2414 | static int tcp_time_to_recover(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2415 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2416 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2417 | __u32 packets_out; |
| 2418 | |
Ryousei Takano | 564262c1 | 2007-10-25 23:03:52 -0700 | [diff] [blame] | 2419 | /* Do not perform any recovery during F-RTO algorithm */ |
Ilpo Järvinen | 52c63f1 | 2007-02-21 23:06:52 -0800 | [diff] [blame] | 2420 | if (tp->frto_counter) |
| 2421 | return 0; |
| 2422 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2423 | /* Trick#1: The loss is proven. */ |
| 2424 | if (tp->lost_out) |
| 2425 | return 1; |
| 2426 | |
| 2427 | /* Not-A-Trick#2 : Classic rule... */ |
Andreas Petlund | ea84e55 | 2009-10-27 03:27:21 +0000 | [diff] [blame] | 2428 | if (tcp_dupack_heuristics(tp) > tp->reordering) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2429 | return 1; |
| 2430 | |
| 2431 | /* Trick#3 : when we use RFC2988 timer restart, fast |
| 2432 | * retransmit can be triggered by timeout of queue head. |
| 2433 | */ |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2434 | if (tcp_is_fack(tp) && tcp_head_timedout(sk)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2435 | return 1; |
| 2436 | |
| 2437 | /* Trick#4: It is still not OK... But will it be useful to delay |
| 2438 | * recovery more? |
| 2439 | */ |
| 2440 | packets_out = tp->packets_out; |
| 2441 | if (packets_out <= tp->reordering && |
| 2442 | tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2443 | !tcp_may_send_now(sk)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2444 | /* We have nothing to send. This connection is limited |
| 2445 | * either by receiver window or by application. |
| 2446 | */ |
| 2447 | return 1; |
| 2448 | } |
| 2449 | |
| 2450 | return 0; |
| 2451 | } |
| 2452 | |
Ilpo Järvinen | 7363a5b | 2009-02-28 04:44:33 +0000 | [diff] [blame] | 2453 | /* New heuristics: it is possible only after we switched to restart timer |
| 2454 | * each time when something is ACKed. Hence, we can detect timed out packets |
| 2455 | * during fast retransmit without falling to slow start. |
| 2456 | * |
| 2457 | * Usefulness of this as is very questionable, since we should know which of |
| 2458 | * the segments is the next to timeout which is relatively expensive to find |
| 2459 | * in general case unless we add some data structure just for that. The |
| 2460 | * current approach certainly won't find the right one too often and when it |
| 2461 | * finally does find _something_ it usually marks large part of the window |
| 2462 | * right away (because a retransmission with a larger timestamp blocks the |
| 2463 | * loop from advancing). -ij |
| 2464 | */ |
| 2465 | static void tcp_timeout_skbs(struct sock *sk) |
| 2466 | { |
| 2467 | struct tcp_sock *tp = tcp_sk(sk); |
| 2468 | struct sk_buff *skb; |
| 2469 | |
| 2470 | if (!tcp_is_fack(tp) || !tcp_head_timedout(sk)) |
| 2471 | return; |
| 2472 | |
| 2473 | skb = tp->scoreboard_skb_hint; |
| 2474 | if (tp->scoreboard_skb_hint == NULL) |
| 2475 | skb = tcp_write_queue_head(sk); |
| 2476 | |
| 2477 | tcp_for_write_queue_from(skb, sk) { |
| 2478 | if (skb == tcp_send_head(sk)) |
| 2479 | break; |
| 2480 | if (!tcp_skb_timedout(sk, skb)) |
| 2481 | break; |
| 2482 | |
| 2483 | tcp_skb_mark_lost(tp, skb); |
| 2484 | } |
| 2485 | |
| 2486 | tp->scoreboard_skb_hint = skb; |
| 2487 | |
| 2488 | tcp_verify_left_out(tp); |
| 2489 | } |
| 2490 | |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2491 | /* Mark head of queue up as lost. With RFC3517 SACK, the packets is |
| 2492 | * is against sacked "cnt", otherwise it's against facked "cnt" |
| 2493 | */ |
Ilpo Järvinen | 1b69d74 | 2008-04-07 22:31:38 -0700 | [diff] [blame] | 2494 | static void tcp_mark_head_lost(struct sock *sk, int packets) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2495 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2496 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2497 | struct sk_buff *skb; |
Ilpo Järvinen | c137f3d | 2008-04-07 22:32:38 -0700 | [diff] [blame] | 2498 | int cnt, oldcnt; |
| 2499 | int err; |
| 2500 | unsigned int mss; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2501 | |
Ilpo Järvinen | 547b792 | 2008-07-25 21:43:18 -0700 | [diff] [blame] | 2502 | WARN_ON(packets > tp->packets_out); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2503 | if (tp->lost_skb_hint) { |
| 2504 | skb = tp->lost_skb_hint; |
| 2505 | cnt = tp->lost_cnt_hint; |
| 2506 | } else { |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2507 | skb = tcp_write_queue_head(sk); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2508 | cnt = 0; |
| 2509 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2510 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2511 | tcp_for_write_queue_from(skb, sk) { |
| 2512 | if (skb == tcp_send_head(sk)) |
| 2513 | break; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2514 | /* TODO: do this better */ |
| 2515 | /* this is not the most efficient way to do this... */ |
| 2516 | tp->lost_skb_hint = skb; |
| 2517 | tp->lost_cnt_hint = cnt; |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2518 | |
Ilpo Järvinen | c137f3d | 2008-04-07 22:32:38 -0700 | [diff] [blame] | 2519 | if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq)) |
| 2520 | break; |
| 2521 | |
| 2522 | oldcnt = cnt; |
Ilpo Järvinen | ad1984e | 2008-01-30 20:06:02 -0800 | [diff] [blame] | 2523 | if (tcp_is_fack(tp) || tcp_is_reno(tp) || |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2524 | (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) |
| 2525 | cnt += tcp_skb_pcount(skb); |
| 2526 | |
Ilpo Järvinen | c137f3d | 2008-04-07 22:32:38 -0700 | [diff] [blame] | 2527 | if (cnt > packets) { |
| 2528 | if (tcp_is_sack(tp) || (oldcnt >= packets)) |
| 2529 | break; |
| 2530 | |
| 2531 | mss = skb_shinfo(skb)->gso_size; |
| 2532 | err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss); |
| 2533 | if (err < 0) |
| 2534 | break; |
| 2535 | cnt = packets; |
| 2536 | } |
| 2537 | |
Ilpo Järvinen | 41ea36e | 2008-09-20 21:19:22 -0700 | [diff] [blame] | 2538 | tcp_skb_mark_lost(tp, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2539 | } |
Ilpo Järvinen | 005903b | 2007-08-09 14:44:16 +0300 | [diff] [blame] | 2540 | tcp_verify_left_out(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2541 | } |
| 2542 | |
| 2543 | /* Account newly detected lost packet(s) */ |
| 2544 | |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2545 | static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2546 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2547 | struct tcp_sock *tp = tcp_sk(sk); |
| 2548 | |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2549 | if (tcp_is_reno(tp)) { |
Ilpo Järvinen | 1b69d74 | 2008-04-07 22:31:38 -0700 | [diff] [blame] | 2550 | tcp_mark_head_lost(sk, 1); |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2551 | } else if (tcp_is_fack(tp)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2552 | int lost = tp->fackets_out - tp->reordering; |
| 2553 | if (lost <= 0) |
| 2554 | lost = 1; |
Ilpo Järvinen | 1b69d74 | 2008-04-07 22:31:38 -0700 | [diff] [blame] | 2555 | tcp_mark_head_lost(sk, lost); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2556 | } else { |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2557 | int sacked_upto = tp->sacked_out - tp->reordering; |
Ilpo Järvinen | 1b69d74 | 2008-04-07 22:31:38 -0700 | [diff] [blame] | 2558 | if (sacked_upto < fast_rexmit) |
| 2559 | sacked_upto = fast_rexmit; |
| 2560 | tcp_mark_head_lost(sk, sacked_upto); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2561 | } |
| 2562 | |
Ilpo Järvinen | 7363a5b | 2009-02-28 04:44:33 +0000 | [diff] [blame] | 2563 | tcp_timeout_skbs(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2564 | } |
| 2565 | |
| 2566 | /* CWND moderation, preventing bursts due to too big ACKs |
| 2567 | * in dubious situations. |
| 2568 | */ |
| 2569 | static inline void tcp_moderate_cwnd(struct tcp_sock *tp) |
| 2570 | { |
| 2571 | tp->snd_cwnd = min(tp->snd_cwnd, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2572 | tcp_packets_in_flight(tp) + tcp_max_burst(tp)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2573 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 2574 | } |
| 2575 | |
Stephen Hemminger | 72dc5b9 | 2006-06-05 17:30:08 -0700 | [diff] [blame] | 2576 | /* Lower bound on congestion window is slow start threshold |
| 2577 | * unless congestion avoidance choice decides to overide it. |
| 2578 | */ |
| 2579 | static inline u32 tcp_cwnd_min(const struct sock *sk) |
| 2580 | { |
| 2581 | const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; |
| 2582 | |
| 2583 | return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh; |
| 2584 | } |
| 2585 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2586 | /* Decrease cwnd each second ack. */ |
Ilpo Järvinen | 1e757f9 | 2007-07-30 19:48:37 -0700 | [diff] [blame] | 2587 | static void tcp_cwnd_down(struct sock *sk, int flag) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2588 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2589 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2590 | int decr = tp->snd_cwnd_cnt + 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2591 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2592 | if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) || |
| 2593 | (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) { |
| 2594 | tp->snd_cwnd_cnt = decr & 1; |
Ilpo Järvinen | 1e757f9 | 2007-07-30 19:48:37 -0700 | [diff] [blame] | 2595 | decr >>= 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2596 | |
Ilpo Järvinen | 1e757f9 | 2007-07-30 19:48:37 -0700 | [diff] [blame] | 2597 | if (decr && tp->snd_cwnd > tcp_cwnd_min(sk)) |
| 2598 | tp->snd_cwnd -= decr; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2599 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2600 | tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1); |
Ilpo Järvinen | 1e757f9 | 2007-07-30 19:48:37 -0700 | [diff] [blame] | 2601 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 2602 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2603 | } |
| 2604 | |
| 2605 | /* Nothing was retransmitted or returned timestamp is less |
| 2606 | * than timestamp of the first retransmission. |
| 2607 | */ |
| 2608 | static inline int tcp_packet_delayed(struct tcp_sock *tp) |
| 2609 | { |
| 2610 | return !tp->retrans_stamp || |
| 2611 | (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && |
Arnd Hannemann | d7ee147 | 2008-04-21 14:46:22 -0700 | [diff] [blame] | 2612 | before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2613 | } |
| 2614 | |
| 2615 | /* Undo procedures. */ |
| 2616 | |
| 2617 | #if FASTRETRANS_DEBUG > 1 |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2618 | static void DBGUNDO(struct sock *sk, const char *msg) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2619 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2620 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2621 | struct inet_sock *inet = inet_sk(sk); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2622 | |
YOSHIFUJI Hideaki | 569508c | 2008-04-14 04:09:36 -0700 | [diff] [blame] | 2623 | if (sk->sk_family == AF_INET) { |
Harvey Harrison | 673d57e | 2008-10-31 00:53:57 -0700 | [diff] [blame] | 2624 | printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n", |
YOSHIFUJI Hideaki | 569508c | 2008-04-14 04:09:36 -0700 | [diff] [blame] | 2625 | msg, |
Harvey Harrison | 673d57e | 2008-10-31 00:53:57 -0700 | [diff] [blame] | 2626 | &inet->daddr, ntohs(inet->dport), |
YOSHIFUJI Hideaki | 569508c | 2008-04-14 04:09:36 -0700 | [diff] [blame] | 2627 | tp->snd_cwnd, tcp_left_out(tp), |
| 2628 | tp->snd_ssthresh, tp->prior_ssthresh, |
| 2629 | tp->packets_out); |
| 2630 | } |
| 2631 | #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) |
| 2632 | else if (sk->sk_family == AF_INET6) { |
| 2633 | struct ipv6_pinfo *np = inet6_sk(sk); |
Harvey Harrison | 5b095d989 | 2008-10-29 12:52:50 -0700 | [diff] [blame] | 2634 | printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n", |
YOSHIFUJI Hideaki | 569508c | 2008-04-14 04:09:36 -0700 | [diff] [blame] | 2635 | msg, |
Harvey Harrison | 0c6ce78 | 2008-10-28 16:09:23 -0700 | [diff] [blame] | 2636 | &np->daddr, ntohs(inet->dport), |
YOSHIFUJI Hideaki | 569508c | 2008-04-14 04:09:36 -0700 | [diff] [blame] | 2637 | tp->snd_cwnd, tcp_left_out(tp), |
| 2638 | tp->snd_ssthresh, tp->prior_ssthresh, |
| 2639 | tp->packets_out); |
| 2640 | } |
| 2641 | #endif |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2642 | } |
| 2643 | #else |
| 2644 | #define DBGUNDO(x...) do { } while (0) |
| 2645 | #endif |
| 2646 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2647 | static void tcp_undo_cwr(struct sock *sk, const int undo) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2648 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2649 | struct tcp_sock *tp = tcp_sk(sk); |
| 2650 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2651 | if (tp->prior_ssthresh) { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2652 | const struct inet_connection_sock *icsk = inet_csk(sk); |
| 2653 | |
| 2654 | if (icsk->icsk_ca_ops->undo_cwnd) |
| 2655 | tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2656 | else |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2657 | tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2658 | |
| 2659 | if (undo && tp->prior_ssthresh > tp->snd_ssthresh) { |
| 2660 | tp->snd_ssthresh = tp->prior_ssthresh; |
| 2661 | TCP_ECN_withdraw_cwr(tp); |
| 2662 | } |
| 2663 | } else { |
| 2664 | tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); |
| 2665 | } |
| 2666 | tcp_moderate_cwnd(tp); |
| 2667 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 2668 | } |
| 2669 | |
| 2670 | static inline int tcp_may_undo(struct tcp_sock *tp) |
| 2671 | { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2672 | return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2673 | } |
| 2674 | |
| 2675 | /* People celebrate: "We love our President!" */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2676 | static int tcp_try_undo_recovery(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2677 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2678 | struct tcp_sock *tp = tcp_sk(sk); |
| 2679 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2680 | if (tcp_may_undo(tp)) { |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 2681 | int mib_idx; |
| 2682 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2683 | /* Happy end! We did not retransmit anything |
| 2684 | * or our original transmission succeeded. |
| 2685 | */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2686 | DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2687 | tcp_undo_cwr(sk, 1); |
| 2688 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 2689 | mib_idx = LINUX_MIB_TCPLOSSUNDO; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2690 | else |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 2691 | mib_idx = LINUX_MIB_TCPFULLUNDO; |
| 2692 | |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 2693 | NET_INC_STATS_BH(sock_net(sk), mib_idx); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2694 | tp->undo_marker = 0; |
| 2695 | } |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 2696 | if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2697 | /* Hold old state until something *above* high_seq |
| 2698 | * is ACKed. For Reno it is MUST to prevent false |
| 2699 | * fast retransmits (RFC2582). SACK TCP is safe. */ |
| 2700 | tcp_moderate_cwnd(tp); |
| 2701 | return 1; |
| 2702 | } |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2703 | tcp_set_ca_state(sk, TCP_CA_Open); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2704 | return 0; |
| 2705 | } |
| 2706 | |
| 2707 | /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2708 | static void tcp_try_undo_dsack(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2709 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2710 | struct tcp_sock *tp = tcp_sk(sk); |
| 2711 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2712 | if (tp->undo_marker && !tp->undo_retrans) { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2713 | DBGUNDO(sk, "D-SACK"); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2714 | tcp_undo_cwr(sk, 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2715 | tp->undo_marker = 0; |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 2716 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2717 | } |
| 2718 | } |
| 2719 | |
Ilpo Järvinen | 77722b1 | 2009-12-08 20:54:11 -0800 | [diff] [blame] | 2720 | /* We can clear retrans_stamp when there are no retransmissions in the |
| 2721 | * window. It would seem that it is trivially available for us in |
| 2722 | * tp->retrans_out, however, that kind of assumptions doesn't consider |
| 2723 | * what will happen if errors occur when sending retransmission for the |
| 2724 | * second time. ...It could the that such segment has only |
| 2725 | * TCPCB_EVER_RETRANS set at the present time. It seems that checking |
| 2726 | * the head skb is enough except for some reneging corner cases that |
| 2727 | * are not worth the effort. |
| 2728 | * |
| 2729 | * Main reason for all this complexity is the fact that connection dying |
| 2730 | * time now depends on the validity of the retrans_stamp, in particular, |
| 2731 | * that successive retransmissions of a segment must not advance |
| 2732 | * retrans_stamp under any conditions. |
| 2733 | */ |
| 2734 | static int tcp_any_retrans_done(struct sock *sk) |
| 2735 | { |
| 2736 | struct tcp_sock *tp = tcp_sk(sk); |
| 2737 | struct sk_buff *skb; |
| 2738 | |
| 2739 | if (tp->retrans_out) |
| 2740 | return 1; |
| 2741 | |
| 2742 | skb = tcp_write_queue_head(sk); |
| 2743 | if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS)) |
| 2744 | return 1; |
| 2745 | |
| 2746 | return 0; |
| 2747 | } |
| 2748 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2749 | /* Undo during fast recovery after partial ACK. */ |
| 2750 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2751 | static int tcp_try_undo_partial(struct sock *sk, int acked) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2752 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2753 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2754 | /* Partial ACK arrived. Force Hoe's retransmit. */ |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2755 | int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2756 | |
| 2757 | if (tcp_may_undo(tp)) { |
| 2758 | /* Plain luck! Hole if filled with delayed |
| 2759 | * packet, rather than with a retransmit. |
| 2760 | */ |
Ilpo Järvinen | 77722b1 | 2009-12-08 20:54:11 -0800 | [diff] [blame] | 2761 | if (!tcp_any_retrans_done(sk)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2762 | tp->retrans_stamp = 0; |
| 2763 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2764 | tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2765 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2766 | DBGUNDO(sk, "Hoe"); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2767 | tcp_undo_cwr(sk, 0); |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 2768 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2769 | |
| 2770 | /* So... Do not make Hoe's retransmit yet. |
| 2771 | * If the first packet was delayed, the rest |
| 2772 | * ones are most probably delayed as well. |
| 2773 | */ |
| 2774 | failed = 0; |
| 2775 | } |
| 2776 | return failed; |
| 2777 | } |
| 2778 | |
| 2779 | /* Undo during loss recovery after partial ACK. */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2780 | static int tcp_try_undo_loss(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2781 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2782 | struct tcp_sock *tp = tcp_sk(sk); |
| 2783 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2784 | if (tcp_may_undo(tp)) { |
| 2785 | struct sk_buff *skb; |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2786 | tcp_for_write_queue(skb, sk) { |
| 2787 | if (skb == tcp_send_head(sk)) |
| 2788 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2789 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; |
| 2790 | } |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2791 | |
Ilpo Järvinen | 5af4ec2 | 2007-09-20 11:30:48 -0700 | [diff] [blame] | 2792 | tcp_clear_all_retrans_hints(tp); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2793 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2794 | DBGUNDO(sk, "partial loss"); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2795 | tp->lost_out = 0; |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2796 | tcp_undo_cwr(sk, 1); |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 2797 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO); |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2798 | inet_csk(sk)->icsk_retransmits = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2799 | tp->undo_marker = 0; |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 2800 | if (tcp_is_sack(tp)) |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2801 | tcp_set_ca_state(sk, TCP_CA_Open); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2802 | return 1; |
| 2803 | } |
| 2804 | return 0; |
| 2805 | } |
| 2806 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2807 | static inline void tcp_complete_cwr(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2808 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2809 | struct tcp_sock *tp = tcp_sk(sk); |
Stephen Hemminger | 317a76f | 2005-06-23 12:19:55 -0700 | [diff] [blame] | 2810 | tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2811 | tp->snd_cwnd_stamp = tcp_time_stamp; |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2812 | tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2813 | } |
| 2814 | |
Ilpo Järvinen | 8aca6cb | 2008-06-04 11:34:22 -0700 | [diff] [blame] | 2815 | static void tcp_try_keep_open(struct sock *sk) |
| 2816 | { |
| 2817 | struct tcp_sock *tp = tcp_sk(sk); |
| 2818 | int state = TCP_CA_Open; |
| 2819 | |
Ilpo Järvinen | 77722b1 | 2009-12-08 20:54:11 -0800 | [diff] [blame] | 2820 | if (tcp_left_out(tp) || tcp_any_retrans_done(sk) || tp->undo_marker) |
Ilpo Järvinen | 8aca6cb | 2008-06-04 11:34:22 -0700 | [diff] [blame] | 2821 | state = TCP_CA_Disorder; |
| 2822 | |
| 2823 | if (inet_csk(sk)->icsk_ca_state != state) { |
| 2824 | tcp_set_ca_state(sk, state); |
| 2825 | tp->high_seq = tp->snd_nxt; |
| 2826 | } |
| 2827 | } |
| 2828 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2829 | static void tcp_try_to_open(struct sock *sk, int flag) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2830 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2831 | struct tcp_sock *tp = tcp_sk(sk); |
| 2832 | |
Ilpo Järvinen | 86426c2 | 2007-08-09 14:45:17 +0300 | [diff] [blame] | 2833 | tcp_verify_left_out(tp); |
| 2834 | |
Ilpo Järvinen | 77722b1 | 2009-12-08 20:54:11 -0800 | [diff] [blame] | 2835 | if (!tp->frto_counter && !tcp_any_retrans_done(sk)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2836 | tp->retrans_stamp = 0; |
| 2837 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2838 | if (flag & FLAG_ECE) |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 2839 | tcp_enter_cwr(sk, 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2840 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2841 | if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { |
Ilpo Järvinen | 8aca6cb | 2008-06-04 11:34:22 -0700 | [diff] [blame] | 2842 | tcp_try_keep_open(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2843 | tcp_moderate_cwnd(tp); |
| 2844 | } else { |
Ilpo Järvinen | 1e757f9 | 2007-07-30 19:48:37 -0700 | [diff] [blame] | 2845 | tcp_cwnd_down(sk, flag); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2846 | } |
| 2847 | } |
| 2848 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2849 | static void tcp_mtup_probe_failed(struct sock *sk) |
| 2850 | { |
| 2851 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 2852 | |
| 2853 | icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; |
| 2854 | icsk->icsk_mtup.probe_size = 0; |
| 2855 | } |
| 2856 | |
Ilpo Järvinen | 72211e9 | 2009-03-14 14:23:04 +0000 | [diff] [blame] | 2857 | static void tcp_mtup_probe_success(struct sock *sk) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2858 | { |
| 2859 | struct tcp_sock *tp = tcp_sk(sk); |
| 2860 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 2861 | |
| 2862 | /* FIXME: breaks with very large cwnd */ |
| 2863 | tp->prior_ssthresh = tcp_current_ssthresh(sk); |
| 2864 | tp->snd_cwnd = tp->snd_cwnd * |
| 2865 | tcp_mss_to_mtu(sk, tp->mss_cache) / |
| 2866 | icsk->icsk_mtup.probe_size; |
| 2867 | tp->snd_cwnd_cnt = 0; |
| 2868 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 2869 | tp->rcv_ssthresh = tcp_current_ssthresh(sk); |
| 2870 | |
| 2871 | icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; |
| 2872 | icsk->icsk_mtup.probe_size = 0; |
| 2873 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); |
| 2874 | } |
| 2875 | |
Ilpo Järvinen | e1aa680 | 2008-11-24 21:11:55 -0800 | [diff] [blame] | 2876 | /* Do a simple retransmit without using the backoff mechanisms in |
| 2877 | * tcp_timer. This is used for path mtu discovery. |
| 2878 | * The socket is already locked here. |
| 2879 | */ |
| 2880 | void tcp_simple_retransmit(struct sock *sk) |
| 2881 | { |
| 2882 | const struct inet_connection_sock *icsk = inet_csk(sk); |
| 2883 | struct tcp_sock *tp = tcp_sk(sk); |
| 2884 | struct sk_buff *skb; |
Ilpo Järvinen | 0c54b85 | 2009-03-14 14:23:05 +0000 | [diff] [blame] | 2885 | unsigned int mss = tcp_current_mss(sk); |
Ilpo Järvinen | e1aa680 | 2008-11-24 21:11:55 -0800 | [diff] [blame] | 2886 | u32 prior_lost = tp->lost_out; |
| 2887 | |
| 2888 | tcp_for_write_queue(skb, sk) { |
| 2889 | if (skb == tcp_send_head(sk)) |
| 2890 | break; |
Ilpo Järvinen | 775ffab | 2008-12-05 22:41:26 -0800 | [diff] [blame] | 2891 | if (tcp_skb_seglen(skb) > mss && |
Ilpo Järvinen | e1aa680 | 2008-11-24 21:11:55 -0800 | [diff] [blame] | 2892 | !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { |
| 2893 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
| 2894 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; |
| 2895 | tp->retrans_out -= tcp_skb_pcount(skb); |
| 2896 | } |
| 2897 | tcp_skb_mark_lost_uncond_verify(tp, skb); |
| 2898 | } |
| 2899 | } |
| 2900 | |
| 2901 | tcp_clear_retrans_hints_partial(tp); |
| 2902 | |
| 2903 | if (prior_lost == tp->lost_out) |
| 2904 | return; |
| 2905 | |
| 2906 | if (tcp_is_reno(tp)) |
| 2907 | tcp_limit_reno_sacked(tp); |
| 2908 | |
| 2909 | tcp_verify_left_out(tp); |
| 2910 | |
| 2911 | /* Don't muck with the congestion window here. |
| 2912 | * Reason is that we do not increase amount of _data_ |
| 2913 | * in network, but units changed and effective |
| 2914 | * cwnd/ssthresh really reduced now. |
| 2915 | */ |
| 2916 | if (icsk->icsk_ca_state != TCP_CA_Loss) { |
| 2917 | tp->high_seq = tp->snd_nxt; |
| 2918 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
| 2919 | tp->prior_ssthresh = 0; |
| 2920 | tp->undo_marker = 0; |
| 2921 | tcp_set_ca_state(sk, TCP_CA_Loss); |
| 2922 | } |
| 2923 | tcp_xmit_retransmit_queue(sk); |
| 2924 | } |
| 2925 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2926 | /* Process an event, which can update packets-in-flight not trivially. |
| 2927 | * Main goal of this function is to calculate new estimate for left_out, |
| 2928 | * taking into account both packets sitting in receiver's buffer and |
| 2929 | * packets lost by network. |
| 2930 | * |
| 2931 | * Besides that it does CWND reduction, when packet loss is detected |
| 2932 | * and changes state of machine. |
| 2933 | * |
| 2934 | * It does _not_ decide what to send, it is made in function |
| 2935 | * tcp_xmit_retransmit_queue(). |
| 2936 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2937 | static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2938 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2939 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2940 | struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2941 | int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP)); |
| 2942 | int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) && |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2943 | (tcp_fackets_out(tp) > tp->reordering)); |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 2944 | int fast_rexmit = 0, mib_idx; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2945 | |
Ilpo Järvinen | 3ccd313 | 2007-12-31 04:43:32 -0800 | [diff] [blame] | 2946 | if (WARN_ON(!tp->packets_out && tp->sacked_out)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2947 | tp->sacked_out = 0; |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 2948 | if (WARN_ON(!tp->sacked_out && tp->fackets_out)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2949 | tp->fackets_out = 0; |
| 2950 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2951 | /* Now state machine starts. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2952 | * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2953 | if (flag & FLAG_ECE) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2954 | tp->prior_ssthresh = 0; |
| 2955 | |
| 2956 | /* B. In all the states check for reneging SACKs. */ |
Ilpo Järvinen | cadbd03 | 2007-12-31 04:49:21 -0800 | [diff] [blame] | 2957 | if (tcp_check_sack_reneging(sk, flag)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2958 | return; |
| 2959 | |
| 2960 | /* C. Process data loss notification, provided it is valid. */ |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2961 | if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) && |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2962 | before(tp->snd_una, tp->high_seq) && |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2963 | icsk->icsk_ca_state != TCP_CA_Open && |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2964 | tp->fackets_out > tp->reordering) { |
Ilpo Järvinen | 1b69d74 | 2008-04-07 22:31:38 -0700 | [diff] [blame] | 2965 | tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering); |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 2966 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2967 | } |
| 2968 | |
Ilpo Järvinen | 005903b | 2007-08-09 14:44:16 +0300 | [diff] [blame] | 2969 | /* D. Check consistency of the current state. */ |
| 2970 | tcp_verify_left_out(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2971 | |
| 2972 | /* E. Check state exit conditions. State can be terminated |
| 2973 | * when high_seq is ACKed. */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2974 | if (icsk->icsk_ca_state == TCP_CA_Open) { |
Ilpo Järvinen | 547b792 | 2008-07-25 21:43:18 -0700 | [diff] [blame] | 2975 | WARN_ON(tp->retrans_out != 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2976 | tp->retrans_stamp = 0; |
| 2977 | } else if (!before(tp->snd_una, tp->high_seq)) { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2978 | switch (icsk->icsk_ca_state) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2979 | case TCP_CA_Loss: |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2980 | icsk->icsk_retransmits = 0; |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2981 | if (tcp_try_undo_recovery(sk)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2982 | return; |
| 2983 | break; |
| 2984 | |
| 2985 | case TCP_CA_CWR: |
| 2986 | /* CWR is to be held something *above* high_seq |
| 2987 | * is ACKed for CWR bit to reach receiver. */ |
| 2988 | if (tp->snd_una != tp->high_seq) { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2989 | tcp_complete_cwr(sk); |
| 2990 | tcp_set_ca_state(sk, TCP_CA_Open); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2991 | } |
| 2992 | break; |
| 2993 | |
| 2994 | case TCP_CA_Disorder: |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2995 | tcp_try_undo_dsack(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2996 | if (!tp->undo_marker || |
| 2997 | /* For SACK case do not Open to allow to undo |
| 2998 | * catching for all duplicate ACKs. */ |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 2999 | tcp_is_reno(tp) || tp->snd_una != tp->high_seq) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3000 | tp->undo_marker = 0; |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3001 | tcp_set_ca_state(sk, TCP_CA_Open); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3002 | } |
| 3003 | break; |
| 3004 | |
| 3005 | case TCP_CA_Recovery: |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 3006 | if (tcp_is_reno(tp)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3007 | tcp_reset_reno_sack(tp); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 3008 | if (tcp_try_undo_recovery(sk)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3009 | return; |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3010 | tcp_complete_cwr(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3011 | break; |
| 3012 | } |
| 3013 | } |
| 3014 | |
| 3015 | /* F. Process state. */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3016 | switch (icsk->icsk_ca_state) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3017 | case TCP_CA_Recovery: |
Ilpo Järvinen | 2e60529 | 2007-08-02 19:46:58 -0700 | [diff] [blame] | 3018 | if (!(flag & FLAG_SND_UNA_ADVANCED)) { |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 3019 | if (tcp_is_reno(tp) && is_dupack) |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3020 | tcp_add_reno_sack(sk); |
Ilpo Järvinen | 1b6d427b | 2007-08-09 14:53:36 +0300 | [diff] [blame] | 3021 | } else |
| 3022 | do_lost = tcp_try_undo_partial(sk, pkts_acked); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3023 | break; |
| 3024 | case TCP_CA_Loss: |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3025 | if (flag & FLAG_DATA_ACKED) |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3026 | icsk->icsk_retransmits = 0; |
Ilpo Järvinen | 882beba | 2008-04-07 22:33:07 -0700 | [diff] [blame] | 3027 | if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED) |
| 3028 | tcp_reset_reno_sack(tp); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 3029 | if (!tcp_try_undo_loss(sk)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3030 | tcp_moderate_cwnd(tp); |
| 3031 | tcp_xmit_retransmit_queue(sk); |
| 3032 | return; |
| 3033 | } |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3034 | if (icsk->icsk_ca_state != TCP_CA_Open) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3035 | return; |
| 3036 | /* Loss is undone; fall through to processing in Open state. */ |
| 3037 | default: |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 3038 | if (tcp_is_reno(tp)) { |
Ilpo Järvinen | 2e60529 | 2007-08-02 19:46:58 -0700 | [diff] [blame] | 3039 | if (flag & FLAG_SND_UNA_ADVANCED) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3040 | tcp_reset_reno_sack(tp); |
| 3041 | if (is_dupack) |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3042 | tcp_add_reno_sack(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3043 | } |
| 3044 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3045 | if (icsk->icsk_ca_state == TCP_CA_Disorder) |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 3046 | tcp_try_undo_dsack(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3047 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 3048 | if (!tcp_time_to_recover(sk)) { |
| 3049 | tcp_try_to_open(sk, flag); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3050 | return; |
| 3051 | } |
| 3052 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 3053 | /* MTU probe failure: don't reduce cwnd */ |
| 3054 | if (icsk->icsk_ca_state < TCP_CA_CWR && |
| 3055 | icsk->icsk_mtup.probe_size && |
John Heffner | 0e7b136 | 2006-03-20 21:32:58 -0800 | [diff] [blame] | 3056 | tp->snd_una == tp->mtu_probe.probe_seq_start) { |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 3057 | tcp_mtup_probe_failed(sk); |
| 3058 | /* Restores the reduction we did in tcp_mtup_probe() */ |
| 3059 | tp->snd_cwnd++; |
| 3060 | tcp_simple_retransmit(sk); |
| 3061 | return; |
| 3062 | } |
| 3063 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3064 | /* Otherwise enter Recovery state */ |
| 3065 | |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 3066 | if (tcp_is_reno(tp)) |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 3067 | mib_idx = LINUX_MIB_TCPRENORECOVERY; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3068 | else |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 3069 | mib_idx = LINUX_MIB_TCPSACKRECOVERY; |
| 3070 | |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 3071 | NET_INC_STATS_BH(sock_net(sk), mib_idx); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3072 | |
| 3073 | tp->high_seq = tp->snd_nxt; |
| 3074 | tp->prior_ssthresh = 0; |
| 3075 | tp->undo_marker = tp->snd_una; |
| 3076 | tp->undo_retrans = tp->retrans_out; |
| 3077 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3078 | if (icsk->icsk_ca_state < TCP_CA_CWR) { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3079 | if (!(flag & FLAG_ECE)) |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3080 | tp->prior_ssthresh = tcp_current_ssthresh(sk); |
| 3081 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3082 | TCP_ECN_queue_cwr(tp); |
| 3083 | } |
| 3084 | |
Stephen Hemminger | 9772efb | 2005-11-10 17:09:53 -0800 | [diff] [blame] | 3085 | tp->bytes_acked = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3086 | tp->snd_cwnd_cnt = 0; |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3087 | tcp_set_ca_state(sk, TCP_CA_Recovery); |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 3088 | fast_rexmit = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3089 | } |
| 3090 | |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 3091 | if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk))) |
| 3092 | tcp_update_scoreboard(sk, fast_rexmit); |
Ilpo Järvinen | 1e757f9 | 2007-07-30 19:48:37 -0700 | [diff] [blame] | 3093 | tcp_cwnd_down(sk, flag); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3094 | tcp_xmit_retransmit_queue(sk); |
| 3095 | } |
| 3096 | |
Ilpo Järvinen | 41834b7 | 2008-12-05 22:43:26 -0800 | [diff] [blame] | 3097 | static void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt) |
| 3098 | { |
| 3099 | tcp_rtt_estimator(sk, seq_rtt); |
| 3100 | tcp_set_rto(sk); |
| 3101 | inet_csk(sk)->icsk_backoff = 0; |
| 3102 | } |
| 3103 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3104 | /* Read draft-ietf-tcplw-high-performance before mucking |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 3105 | * with this code. (Supersedes RFC1323) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3106 | */ |
Stephen Hemminger | 2d2abba | 2005-11-10 16:56:12 -0800 | [diff] [blame] | 3107 | static void tcp_ack_saw_tstamp(struct sock *sk, int flag) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3108 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3109 | /* RTTM Rule: A TSecr value received in a segment is used to |
| 3110 | * update the averaged RTT measurement only if the segment |
| 3111 | * acknowledges some new data, i.e., only if it advances the |
| 3112 | * left edge of the send window. |
| 3113 | * |
| 3114 | * See draft-ietf-tcplw-high-performance-00, section 3.3. |
| 3115 | * 1998/04/10 Andrey V. Savochkin <saw@msu.ru> |
| 3116 | * |
| 3117 | * Changed: reset backoff as soon as we see the first valid sample. |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 3118 | * If we do not, we get strongly overestimated rto. With timestamps |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3119 | * samples are accepted even from very old segments: f.e., when rtt=1 |
| 3120 | * increases to 8, we retransmit 5 times and after 8 seconds delayed |
| 3121 | * answer arrives rto becomes 120 seconds! If at least one of segments |
| 3122 | * in window is lost... Voila. --ANK (010210) |
| 3123 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3124 | struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 41834b7 | 2008-12-05 22:43:26 -0800 | [diff] [blame] | 3125 | |
| 3126 | tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3127 | } |
| 3128 | |
Stephen Hemminger | 2d2abba | 2005-11-10 16:56:12 -0800 | [diff] [blame] | 3129 | static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3130 | { |
| 3131 | /* We don't have a timestamp. Can only use |
| 3132 | * packets that are not retransmitted to determine |
| 3133 | * rtt estimates. Also, we must not reset the |
| 3134 | * backoff for rto until we get a non-retransmitted |
| 3135 | * packet. This allows us to deal with a situation |
| 3136 | * where the network delay has increased suddenly. |
| 3137 | * I.e. Karn's algorithm. (SIGCOMM '87, p5.) |
| 3138 | */ |
| 3139 | |
| 3140 | if (flag & FLAG_RETRANS_DATA_ACKED) |
| 3141 | return; |
| 3142 | |
Ilpo Järvinen | 41834b7 | 2008-12-05 22:43:26 -0800 | [diff] [blame] | 3143 | tcp_valid_rtt_meas(sk, seq_rtt); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3144 | } |
| 3145 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3146 | static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, |
Stephen Hemminger | 2d2abba | 2005-11-10 16:56:12 -0800 | [diff] [blame] | 3147 | const s32 seq_rtt) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3148 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3149 | const struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3150 | /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ |
| 3151 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) |
Stephen Hemminger | 2d2abba | 2005-11-10 16:56:12 -0800 | [diff] [blame] | 3152 | tcp_ack_saw_tstamp(sk, flag); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3153 | else if (seq_rtt >= 0) |
Stephen Hemminger | 2d2abba | 2005-11-10 16:56:12 -0800 | [diff] [blame] | 3154 | tcp_ack_no_tstamp(sk, seq_rtt, flag); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3155 | } |
| 3156 | |
Ilpo Järvinen | c3a05c6 | 2007-12-02 00:47:59 +0200 | [diff] [blame] | 3157 | static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3158 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3159 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Ilpo Järvinen | c3a05c6 | 2007-12-02 00:47:59 +0200 | [diff] [blame] | 3160 | icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3161 | tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3162 | } |
| 3163 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3164 | /* Restart timer after forward progress on connection. |
| 3165 | * RFC2988 recommends to restart timer to now+rto. |
| 3166 | */ |
Ilpo Järvinen | 6728e7d | 2007-08-24 22:53:26 -0700 | [diff] [blame] | 3167 | static void tcp_rearm_rto(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3168 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 3169 | struct tcp_sock *tp = tcp_sk(sk); |
| 3170 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3171 | if (!tp->packets_out) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3172 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3173 | } else { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3174 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
| 3175 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3176 | } |
| 3177 | } |
| 3178 | |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3179 | /* If we get here, the whole TSO packet has not been acked. */ |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 3180 | static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3181 | { |
| 3182 | struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3183 | u32 packets_acked; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3184 | |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3185 | BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3186 | |
| 3187 | packets_acked = tcp_skb_pcount(skb); |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3188 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3189 | return 0; |
| 3190 | packets_acked -= tcp_skb_pcount(skb); |
| 3191 | |
| 3192 | if (packets_acked) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3193 | BUG_ON(tcp_skb_pcount(skb) == 0); |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3194 | BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3195 | } |
| 3196 | |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 3197 | return packets_acked; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3198 | } |
| 3199 | |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3200 | /* Remove acknowledged frames from the retransmission queue. If our packet |
| 3201 | * is before the ack sequence we can discard it as it's confirmed to have |
| 3202 | * arrived at the other end. |
| 3203 | */ |
Ilpo Järvinen | 33f5f57 | 2008-10-07 14:43:06 -0700 | [diff] [blame] | 3204 | static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets, |
| 3205 | u32 prior_snd_una) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3206 | { |
| 3207 | struct tcp_sock *tp = tcp_sk(sk); |
Stephen Hemminger | 2d2abba | 2005-11-10 16:56:12 -0800 | [diff] [blame] | 3208 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3209 | struct sk_buff *skb; |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3210 | u32 now = tcp_time_stamp; |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 3211 | int fully_acked = 1; |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3212 | int flag = 0; |
Ilpo Järvinen | 7201883 | 2007-12-30 04:37:55 -0800 | [diff] [blame] | 3213 | u32 pkts_acked = 0; |
Ilpo Järvinen | c7caf8d | 2007-11-10 21:22:18 -0800 | [diff] [blame] | 3214 | u32 reord = tp->packets_out; |
Ilpo Järvinen | 90638a0 | 2008-09-20 21:25:52 -0700 | [diff] [blame] | 3215 | u32 prior_sacked = tp->sacked_out; |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3216 | s32 seq_rtt = -1; |
Gavin McCullagh | 2072c22 | 2007-12-29 19:11:21 -0800 | [diff] [blame] | 3217 | s32 ca_seq_rtt = -1; |
Ilpo Järvinen | b9ce204 | 2007-06-15 15:08:43 -0700 | [diff] [blame] | 3218 | ktime_t last_ackt = net_invalid_timestamp(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3219 | |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3220 | while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) { |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 3221 | struct tcp_skb_cb *scb = TCP_SKB_CB(skb); |
Ilpo Järvinen | 7201883 | 2007-12-30 04:37:55 -0800 | [diff] [blame] | 3222 | u32 acked_pcount; |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3223 | u8 sacked = scb->sacked; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3224 | |
Gavin McCullagh | 2072c22 | 2007-12-29 19:11:21 -0800 | [diff] [blame] | 3225 | /* Determine how many packets and what bytes were acked, tso and else */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3226 | if (after(scb->end_seq, tp->snd_una)) { |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 3227 | if (tcp_skb_pcount(skb) == 1 || |
| 3228 | !after(tp->snd_una, scb->seq)) |
| 3229 | break; |
| 3230 | |
Ilpo Järvinen | 7201883 | 2007-12-30 04:37:55 -0800 | [diff] [blame] | 3231 | acked_pcount = tcp_tso_acked(sk, skb); |
| 3232 | if (!acked_pcount) |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 3233 | break; |
| 3234 | |
| 3235 | fully_acked = 0; |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 3236 | } else { |
Ilpo Järvinen | 7201883 | 2007-12-30 04:37:55 -0800 | [diff] [blame] | 3237 | acked_pcount = tcp_skb_pcount(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3238 | } |
| 3239 | |
Ilpo Järvinen | 89d478f | 2007-12-30 04:35:27 -0800 | [diff] [blame] | 3240 | if (sacked & TCPCB_RETRANS) { |
| 3241 | if (sacked & TCPCB_SACKED_RETRANS) |
Ilpo Järvinen | 7201883 | 2007-12-30 04:37:55 -0800 | [diff] [blame] | 3242 | tp->retrans_out -= acked_pcount; |
Ilpo Järvinen | 89d478f | 2007-12-30 04:35:27 -0800 | [diff] [blame] | 3243 | flag |= FLAG_RETRANS_DATA_ACKED; |
| 3244 | ca_seq_rtt = -1; |
| 3245 | seq_rtt = -1; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3246 | if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1)) |
Ilpo Järvinen | 89d478f | 2007-12-30 04:35:27 -0800 | [diff] [blame] | 3247 | flag |= FLAG_NONHEAD_RETRANS_ACKED; |
Ilpo Järvinen | c7caf8d | 2007-11-10 21:22:18 -0800 | [diff] [blame] | 3248 | } else { |
Gavin McCullagh | 2072c22 | 2007-12-29 19:11:21 -0800 | [diff] [blame] | 3249 | ca_seq_rtt = now - scb->when; |
| 3250 | last_ackt = skb->tstamp; |
Ilpo Järvinen | c7caf8d | 2007-11-10 21:22:18 -0800 | [diff] [blame] | 3251 | if (seq_rtt < 0) { |
Gavin McCullagh | 2072c22 | 2007-12-29 19:11:21 -0800 | [diff] [blame] | 3252 | seq_rtt = ca_seq_rtt; |
Ilpo Järvinen | c7caf8d | 2007-11-10 21:22:18 -0800 | [diff] [blame] | 3253 | } |
Ilpo Järvinen | 89d478f | 2007-12-30 04:35:27 -0800 | [diff] [blame] | 3254 | if (!(sacked & TCPCB_SACKED_ACKED)) |
Ilpo Järvinen | 7201883 | 2007-12-30 04:37:55 -0800 | [diff] [blame] | 3255 | reord = min(pkts_acked, reord); |
Stephen Hemminger | 2d2abba | 2005-11-10 16:56:12 -0800 | [diff] [blame] | 3256 | } |
Ilpo Järvinen | 89d478f | 2007-12-30 04:35:27 -0800 | [diff] [blame] | 3257 | |
| 3258 | if (sacked & TCPCB_SACKED_ACKED) |
Ilpo Järvinen | 7201883 | 2007-12-30 04:37:55 -0800 | [diff] [blame] | 3259 | tp->sacked_out -= acked_pcount; |
Ilpo Järvinen | 89d478f | 2007-12-30 04:35:27 -0800 | [diff] [blame] | 3260 | if (sacked & TCPCB_LOST) |
Ilpo Järvinen | 7201883 | 2007-12-30 04:37:55 -0800 | [diff] [blame] | 3261 | tp->lost_out -= acked_pcount; |
Ilpo Järvinen | 89d478f | 2007-12-30 04:35:27 -0800 | [diff] [blame] | 3262 | |
Ilpo Järvinen | 7201883 | 2007-12-30 04:37:55 -0800 | [diff] [blame] | 3263 | tp->packets_out -= acked_pcount; |
| 3264 | pkts_acked += acked_pcount; |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 3265 | |
Ilpo Järvinen | 009a2e3 | 2007-09-20 11:34:38 -0700 | [diff] [blame] | 3266 | /* Initial outgoing SYN's get put onto the write_queue |
| 3267 | * just like anything else we transmit. It is not |
| 3268 | * true data, and if we misinform our callers that |
| 3269 | * this ACK acks real data, we will erroneously exit |
| 3270 | * connection startup slow start one packet too |
| 3271 | * quickly. This is severely frowned upon behavior. |
| 3272 | */ |
| 3273 | if (!(scb->flags & TCPCB_FLAG_SYN)) { |
| 3274 | flag |= FLAG_DATA_ACKED; |
| 3275 | } else { |
| 3276 | flag |= FLAG_SYN_ACKED; |
| 3277 | tp->retrans_stamp = 0; |
| 3278 | } |
| 3279 | |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 3280 | if (!fully_acked) |
| 3281 | break; |
| 3282 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 3283 | tcp_unlink_write_queue(skb, sk); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 3284 | sk_wmem_free_skb(sk, skb); |
Ilpo Järvinen | 90638a0 | 2008-09-20 21:25:52 -0700 | [diff] [blame] | 3285 | tp->scoreboard_skb_hint = NULL; |
Ilpo Järvinen | ef9da47 | 2008-09-20 21:25:15 -0700 | [diff] [blame] | 3286 | if (skb == tp->retransmit_skb_hint) |
| 3287 | tp->retransmit_skb_hint = NULL; |
Ilpo Järvinen | 90638a0 | 2008-09-20 21:25:52 -0700 | [diff] [blame] | 3288 | if (skb == tp->lost_skb_hint) |
| 3289 | tp->lost_skb_hint = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3290 | } |
| 3291 | |
Ilpo Järvinen | 33f5f57 | 2008-10-07 14:43:06 -0700 | [diff] [blame] | 3292 | if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una))) |
| 3293 | tp->snd_up = tp->snd_una; |
| 3294 | |
Ilpo Järvinen | cadbd03 | 2007-12-31 04:49:21 -0800 | [diff] [blame] | 3295 | if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) |
| 3296 | flag |= FLAG_SACK_RENEGING; |
| 3297 | |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3298 | if (flag & FLAG_ACKED) { |
Stephen Hemminger | 164891a | 2007-04-23 22:26:16 -0700 | [diff] [blame] | 3299 | const struct tcp_congestion_ops *ca_ops |
| 3300 | = inet_csk(sk)->icsk_ca_ops; |
| 3301 | |
Ilpo Järvinen | 72211e9 | 2009-03-14 14:23:04 +0000 | [diff] [blame] | 3302 | if (unlikely(icsk->icsk_mtup.probe_size && |
| 3303 | !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) { |
| 3304 | tcp_mtup_probe_success(sk); |
| 3305 | } |
| 3306 | |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3307 | tcp_ack_update_rtt(sk, flag, seq_rtt); |
Ilpo Järvinen | 6728e7d | 2007-08-24 22:53:26 -0700 | [diff] [blame] | 3308 | tcp_rearm_rto(sk); |
Stephen Hemminger | 317a76f | 2005-06-23 12:19:55 -0700 | [diff] [blame] | 3309 | |
Ilpo Järvinen | c7caf8d | 2007-11-10 21:22:18 -0800 | [diff] [blame] | 3310 | if (tcp_is_reno(tp)) { |
| 3311 | tcp_remove_reno_sacks(sk, pkts_acked); |
| 3312 | } else { |
Ilpo Järvinen | 59a08cb | 2009-02-28 04:44:28 +0000 | [diff] [blame] | 3313 | int delta; |
| 3314 | |
Ilpo Järvinen | c7caf8d | 2007-11-10 21:22:18 -0800 | [diff] [blame] | 3315 | /* Non-retransmitted hole got filled? That's reordering */ |
| 3316 | if (reord < prior_fackets) |
| 3317 | tcp_update_reordering(sk, tp->fackets_out - reord, 0); |
Ilpo Järvinen | 90638a0 | 2008-09-20 21:25:52 -0700 | [diff] [blame] | 3318 | |
Ilpo Järvinen | 59a08cb | 2009-02-28 04:44:28 +0000 | [diff] [blame] | 3319 | delta = tcp_is_fack(tp) ? pkts_acked : |
| 3320 | prior_sacked - tp->sacked_out; |
| 3321 | tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta); |
Ilpo Järvinen | c7caf8d | 2007-11-10 21:22:18 -0800 | [diff] [blame] | 3322 | } |
| 3323 | |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 3324 | tp->fackets_out -= min(pkts_acked, tp->fackets_out); |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 3325 | |
Stephen Hemminger | 30cfd0b | 2007-07-25 23:49:34 -0700 | [diff] [blame] | 3326 | if (ca_ops->pkts_acked) { |
| 3327 | s32 rtt_us = -1; |
Ilpo Järvinen | b9ce204 | 2007-06-15 15:08:43 -0700 | [diff] [blame] | 3328 | |
Stephen Hemminger | 30cfd0b | 2007-07-25 23:49:34 -0700 | [diff] [blame] | 3329 | /* Is the ACK triggering packet unambiguous? */ |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3330 | if (!(flag & FLAG_RETRANS_DATA_ACKED)) { |
Stephen Hemminger | 30cfd0b | 2007-07-25 23:49:34 -0700 | [diff] [blame] | 3331 | /* High resolution needed and available? */ |
| 3332 | if (ca_ops->flags & TCP_CONG_RTT_STAMP && |
| 3333 | !ktime_equal(last_ackt, |
| 3334 | net_invalid_timestamp())) |
| 3335 | rtt_us = ktime_us_delta(ktime_get_real(), |
| 3336 | last_ackt); |
Gavin McCullagh | 2072c22 | 2007-12-29 19:11:21 -0800 | [diff] [blame] | 3337 | else if (ca_seq_rtt > 0) |
| 3338 | rtt_us = jiffies_to_usecs(ca_seq_rtt); |
Stephen Hemminger | 30cfd0b | 2007-07-25 23:49:34 -0700 | [diff] [blame] | 3339 | } |
| 3340 | |
| 3341 | ca_ops->pkts_acked(sk, pkts_acked, rtt_us); |
| 3342 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3343 | } |
| 3344 | |
| 3345 | #if FASTRETRANS_DEBUG > 0 |
Ilpo Järvinen | 547b792 | 2008-07-25 21:43:18 -0700 | [diff] [blame] | 3346 | WARN_ON((int)tp->sacked_out < 0); |
| 3347 | WARN_ON((int)tp->lost_out < 0); |
| 3348 | WARN_ON((int)tp->retrans_out < 0); |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 3349 | if (!tp->packets_out && tcp_is_sack(tp)) { |
Stephen Hemminger | cfcabdc | 2007-10-09 01:59:42 -0700 | [diff] [blame] | 3350 | icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3351 | if (tp->lost_out) { |
| 3352 | printk(KERN_DEBUG "Leak l=%u %d\n", |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3353 | tp->lost_out, icsk->icsk_ca_state); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3354 | tp->lost_out = 0; |
| 3355 | } |
| 3356 | if (tp->sacked_out) { |
| 3357 | printk(KERN_DEBUG "Leak s=%u %d\n", |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3358 | tp->sacked_out, icsk->icsk_ca_state); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3359 | tp->sacked_out = 0; |
| 3360 | } |
| 3361 | if (tp->retrans_out) { |
| 3362 | printk(KERN_DEBUG "Leak r=%u %d\n", |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3363 | tp->retrans_out, icsk->icsk_ca_state); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3364 | tp->retrans_out = 0; |
| 3365 | } |
| 3366 | } |
| 3367 | #endif |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3368 | return flag; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3369 | } |
| 3370 | |
| 3371 | static void tcp_ack_probe(struct sock *sk) |
| 3372 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3373 | const struct tcp_sock *tp = tcp_sk(sk); |
| 3374 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3375 | |
| 3376 | /* Was it a usable window open? */ |
| 3377 | |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 3378 | if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3379 | icsk->icsk_backoff = 0; |
| 3380 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3381 | /* Socket must be waked up by subsequent tcp_data_snd_check(). |
| 3382 | * This function is not for random using! |
| 3383 | */ |
| 3384 | } else { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3385 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 3386 | min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), |
| 3387 | TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3388 | } |
| 3389 | } |
| 3390 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3391 | static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3392 | { |
| 3393 | return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3394 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3395 | } |
| 3396 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3397 | static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3398 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3399 | const struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3400 | return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3401 | !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3402 | } |
| 3403 | |
| 3404 | /* Check that window update is acceptable. |
| 3405 | * The function assumes that snd_una<=ack<=snd_next. |
| 3406 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3407 | static inline int tcp_may_update_window(const struct tcp_sock *tp, |
| 3408 | const u32 ack, const u32 ack_seq, |
| 3409 | const u32 nwin) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3410 | { |
| 3411 | return (after(ack, tp->snd_una) || |
| 3412 | after(ack_seq, tp->snd_wl1) || |
| 3413 | (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd)); |
| 3414 | } |
| 3415 | |
| 3416 | /* Update our send window. |
| 3417 | * |
| 3418 | * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 |
| 3419 | * and in FreeBSD. NetBSD's one is even worse.) is wrong. |
| 3420 | */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 3421 | static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack, |
| 3422 | u32 ack_seq) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3423 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 3424 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3425 | int flag = 0; |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 3426 | u32 nwin = ntohs(tcp_hdr(skb)->window); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3427 | |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 3428 | if (likely(!tcp_hdr(skb)->syn)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3429 | nwin <<= tp->rx_opt.snd_wscale; |
| 3430 | |
| 3431 | if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { |
| 3432 | flag |= FLAG_WIN_UPDATE; |
Hantzis Fotis | ee7537b | 2009-03-02 22:42:02 -0800 | [diff] [blame] | 3433 | tcp_update_wl(tp, ack_seq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3434 | |
| 3435 | if (tp->snd_wnd != nwin) { |
| 3436 | tp->snd_wnd = nwin; |
| 3437 | |
| 3438 | /* Note, it is the only place, where |
| 3439 | * fast path is recovered for sending TCP. |
| 3440 | */ |
Herbert Xu | 2ad4106 | 2005-10-27 18:47:46 +1000 | [diff] [blame] | 3441 | tp->pred_flags = 0; |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 3442 | tcp_fast_path_check(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3443 | |
| 3444 | if (nwin > tp->max_window) { |
| 3445 | tp->max_window = nwin; |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 3446 | tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3447 | } |
| 3448 | } |
| 3449 | } |
| 3450 | |
| 3451 | tp->snd_una = ack; |
| 3452 | |
| 3453 | return flag; |
| 3454 | } |
| 3455 | |
Ilpo Järvinen | 9ead9a1 | 2007-02-21 22:56:19 -0800 | [diff] [blame] | 3456 | /* A very conservative spurious RTO response algorithm: reduce cwnd and |
| 3457 | * continue in congestion avoidance. |
| 3458 | */ |
| 3459 | static void tcp_conservative_spur_to_response(struct tcp_sock *tp) |
| 3460 | { |
| 3461 | tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); |
Ilpo Järvinen | aa8b6a7 | 2007-02-21 23:06:03 -0800 | [diff] [blame] | 3462 | tp->snd_cwnd_cnt = 0; |
Ilpo Järvinen | 16e9068 | 2007-10-11 17:32:31 -0700 | [diff] [blame] | 3463 | tp->bytes_acked = 0; |
Ilpo Järvinen | 4632365 | 2007-05-19 13:56:23 -0700 | [diff] [blame] | 3464 | TCP_ECN_queue_cwr(tp); |
Ilpo Järvinen | 9ead9a1 | 2007-02-21 22:56:19 -0800 | [diff] [blame] | 3465 | tcp_moderate_cwnd(tp); |
| 3466 | } |
| 3467 | |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 3468 | /* A conservative spurious RTO response algorithm: reduce cwnd using |
| 3469 | * rate halving and continue in congestion avoidance. |
| 3470 | */ |
| 3471 | static void tcp_ratehalving_spur_to_response(struct sock *sk) |
| 3472 | { |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 3473 | tcp_enter_cwr(sk, 0); |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 3474 | } |
| 3475 | |
Ilpo Järvinen | e317f6f | 2007-03-02 13:34:19 -0800 | [diff] [blame] | 3476 | static void tcp_undo_spur_to_response(struct sock *sk, int flag) |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 3477 | { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3478 | if (flag & FLAG_ECE) |
Ilpo Järvinen | e317f6f | 2007-03-02 13:34:19 -0800 | [diff] [blame] | 3479 | tcp_ratehalving_spur_to_response(sk); |
| 3480 | else |
| 3481 | tcp_undo_cwr(sk, 1); |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 3482 | } |
| 3483 | |
Ilpo Järvinen | 30935cf | 2007-02-21 23:01:36 -0800 | [diff] [blame] | 3484 | /* F-RTO spurious RTO detection algorithm (RFC4138) |
| 3485 | * |
Ilpo Järvinen | 6408d20 | 2007-02-21 23:04:11 -0800 | [diff] [blame] | 3486 | * F-RTO affects during two new ACKs following RTO (well, almost, see inline |
| 3487 | * comments). State (ACK number) is kept in frto_counter. When ACK advances |
| 3488 | * window (but not to or beyond highest sequence sent before RTO): |
Ilpo Järvinen | 30935cf | 2007-02-21 23:01:36 -0800 | [diff] [blame] | 3489 | * On First ACK, send two new segments out. |
| 3490 | * On Second ACK, RTO was likely spurious. Do spurious response (response |
| 3491 | * algorithm is not part of the F-RTO detection algorithm |
| 3492 | * given in RFC4138 but can be selected separately). |
| 3493 | * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss |
Ilpo Järvinen | d551e45 | 2007-04-30 00:42:20 -0700 | [diff] [blame] | 3494 | * and TCP falls back to conventional RTO recovery. F-RTO allows overriding |
| 3495 | * of Nagle, this is done using frto_counter states 2 and 3, when a new data |
| 3496 | * segment of any size sent during F-RTO, state 2 is upgraded to 3. |
Ilpo Järvinen | 30935cf | 2007-02-21 23:01:36 -0800 | [diff] [blame] | 3497 | * |
| 3498 | * Rationale: if the RTO was spurious, new ACKs should arrive from the |
| 3499 | * original window even after we transmit two new data segments. |
| 3500 | * |
Ilpo Järvinen | 4dc2665 | 2007-02-21 23:16:11 -0800 | [diff] [blame] | 3501 | * SACK version: |
| 3502 | * on first step, wait until first cumulative ACK arrives, then move to |
| 3503 | * the second step. In second step, the next ACK decides. |
| 3504 | * |
Ilpo Järvinen | 30935cf | 2007-02-21 23:01:36 -0800 | [diff] [blame] | 3505 | * F-RTO is implemented (mainly) in four functions: |
| 3506 | * - tcp_use_frto() is used to determine if TCP is can use F-RTO |
| 3507 | * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is |
| 3508 | * called when tcp_use_frto() showed green light |
| 3509 | * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm |
| 3510 | * - tcp_enter_frto_loss() is called if there is not enough evidence |
| 3511 | * to prove that the RTO is indeed spurious. It transfers the control |
| 3512 | * from F-RTO to the conventional RTO recovery |
| 3513 | */ |
Ilpo Järvinen | 2e60529 | 2007-08-02 19:46:58 -0700 | [diff] [blame] | 3514 | static int tcp_process_frto(struct sock *sk, int flag) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3515 | { |
| 3516 | struct tcp_sock *tp = tcp_sk(sk); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 3517 | |
Ilpo Järvinen | 005903b | 2007-08-09 14:44:16 +0300 | [diff] [blame] | 3518 | tcp_verify_left_out(tp); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 3519 | |
Ilpo Järvinen | 7487c48 | 2007-02-21 23:02:30 -0800 | [diff] [blame] | 3520 | /* Duplicate the behavior from Loss state (fastretrans_alert) */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3521 | if (flag & FLAG_DATA_ACKED) |
Ilpo Järvinen | 7487c48 | 2007-02-21 23:02:30 -0800 | [diff] [blame] | 3522 | inet_csk(sk)->icsk_retransmits = 0; |
| 3523 | |
Ilpo Järvinen | 009a2e3 | 2007-09-20 11:34:38 -0700 | [diff] [blame] | 3524 | if ((flag & FLAG_NONHEAD_RETRANS_ACKED) || |
| 3525 | ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED))) |
| 3526 | tp->undo_marker = 0; |
| 3527 | |
Ilpo Järvinen | 95c4922 | 2007-02-21 23:05:18 -0800 | [diff] [blame] | 3528 | if (!before(tp->snd_una, tp->frto_highmark)) { |
Ilpo Järvinen | d551e45 | 2007-04-30 00:42:20 -0700 | [diff] [blame] | 3529 | tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag); |
Ilpo Järvinen | 7c9a4a5 | 2007-02-21 23:08:34 -0800 | [diff] [blame] | 3530 | return 1; |
Ilpo Järvinen | 95c4922 | 2007-02-21 23:05:18 -0800 | [diff] [blame] | 3531 | } |
| 3532 | |
Ilpo Järvinen | 62ab222 | 2008-05-08 01:09:11 -0700 | [diff] [blame] | 3533 | if (!tcp_is_sackfrto(tp)) { |
Ilpo Järvinen | 4dc2665 | 2007-02-21 23:16:11 -0800 | [diff] [blame] | 3534 | /* RFC4138 shortcoming in step 2; should also have case c): |
| 3535 | * ACK isn't duplicate nor advances window, e.g., opposite dir |
| 3536 | * data, winupdate |
| 3537 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3538 | if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP)) |
Ilpo Järvinen | 4dc2665 | 2007-02-21 23:16:11 -0800 | [diff] [blame] | 3539 | return 1; |
Ilpo Järvinen | 6408d20 | 2007-02-21 23:04:11 -0800 | [diff] [blame] | 3540 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3541 | if (!(flag & FLAG_DATA_ACKED)) { |
Ilpo Järvinen | 4dc2665 | 2007-02-21 23:16:11 -0800 | [diff] [blame] | 3542 | tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3), |
| 3543 | flag); |
| 3544 | return 1; |
| 3545 | } |
| 3546 | } else { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3547 | if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) { |
Ilpo Järvinen | 4dc2665 | 2007-02-21 23:16:11 -0800 | [diff] [blame] | 3548 | /* Prevent sending of new data. */ |
| 3549 | tp->snd_cwnd = min(tp->snd_cwnd, |
| 3550 | tcp_packets_in_flight(tp)); |
| 3551 | return 1; |
| 3552 | } |
| 3553 | |
Ilpo Järvinen | d551e45 | 2007-04-30 00:42:20 -0700 | [diff] [blame] | 3554 | if ((tp->frto_counter >= 2) && |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3555 | (!(flag & FLAG_FORWARD_PROGRESS) || |
| 3556 | ((flag & FLAG_DATA_SACKED) && |
| 3557 | !(flag & FLAG_ONLY_ORIG_SACKED)))) { |
Ilpo Järvinen | 4dc2665 | 2007-02-21 23:16:11 -0800 | [diff] [blame] | 3558 | /* RFC4138 shortcoming (see comment above) */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3559 | if (!(flag & FLAG_FORWARD_PROGRESS) && |
| 3560 | (flag & FLAG_NOT_DUP)) |
Ilpo Järvinen | 4dc2665 | 2007-02-21 23:16:11 -0800 | [diff] [blame] | 3561 | return 1; |
| 3562 | |
| 3563 | tcp_enter_frto_loss(sk, 3, flag); |
| 3564 | return 1; |
| 3565 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3566 | } |
| 3567 | |
| 3568 | if (tp->frto_counter == 1) { |
Ilpo Järvinen | 3e6f049 | 2007-12-05 02:20:21 -0800 | [diff] [blame] | 3569 | /* tcp_may_send_now needs to see updated state */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3570 | tp->snd_cwnd = tcp_packets_in_flight(tp) + 2; |
Ilpo Järvinen | 94d0ea7 | 2007-02-21 23:07:27 -0800 | [diff] [blame] | 3571 | tp->frto_counter = 2; |
Ilpo Järvinen | 3e6f049 | 2007-12-05 02:20:21 -0800 | [diff] [blame] | 3572 | |
| 3573 | if (!tcp_may_send_now(sk)) |
| 3574 | tcp_enter_frto_loss(sk, 2, flag); |
| 3575 | |
Ilpo Järvinen | 7c9a4a5 | 2007-02-21 23:08:34 -0800 | [diff] [blame] | 3576 | return 1; |
Ilpo Järvinen | d551e45 | 2007-04-30 00:42:20 -0700 | [diff] [blame] | 3577 | } else { |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 3578 | switch (sysctl_tcp_frto_response) { |
| 3579 | case 2: |
Ilpo Järvinen | e317f6f | 2007-03-02 13:34:19 -0800 | [diff] [blame] | 3580 | tcp_undo_spur_to_response(sk, flag); |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 3581 | break; |
| 3582 | case 1: |
| 3583 | tcp_conservative_spur_to_response(tp); |
| 3584 | break; |
| 3585 | default: |
| 3586 | tcp_ratehalving_spur_to_response(sk); |
| 3587 | break; |
Stephen Hemminger | 3ff50b7 | 2007-04-20 17:09:22 -0700 | [diff] [blame] | 3588 | } |
Ilpo Järvinen | 94d0ea7 | 2007-02-21 23:07:27 -0800 | [diff] [blame] | 3589 | tp->frto_counter = 0; |
Ilpo Järvinen | 009a2e3 | 2007-09-20 11:34:38 -0700 | [diff] [blame] | 3590 | tp->undo_marker = 0; |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 3591 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3592 | } |
Ilpo Järvinen | 7c9a4a5 | 2007-02-21 23:08:34 -0800 | [diff] [blame] | 3593 | return 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3594 | } |
| 3595 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3596 | /* This routine deals with incoming acks, but not outgoing ones. */ |
| 3597 | static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag) |
| 3598 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3599 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3600 | struct tcp_sock *tp = tcp_sk(sk); |
| 3601 | u32 prior_snd_una = tp->snd_una; |
| 3602 | u32 ack_seq = TCP_SKB_CB(skb)->seq; |
| 3603 | u32 ack = TCP_SKB_CB(skb)->ack_seq; |
| 3604 | u32 prior_in_flight; |
Ilpo Järvinen | c7caf8d | 2007-11-10 21:22:18 -0800 | [diff] [blame] | 3605 | u32 prior_fackets; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3606 | int prior_packets; |
Ilpo Järvinen | 7c9a4a5 | 2007-02-21 23:08:34 -0800 | [diff] [blame] | 3607 | int frto_cwnd = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3608 | |
John Dykstra | 96e0bf4 | 2009-03-22 21:49:57 -0700 | [diff] [blame] | 3609 | /* If the ack is older than previous acks |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3610 | * then we can probably ignore it. |
| 3611 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3612 | if (before(ack, prior_snd_una)) |
| 3613 | goto old_ack; |
| 3614 | |
John Dykstra | 96e0bf4 | 2009-03-22 21:49:57 -0700 | [diff] [blame] | 3615 | /* If the ack includes data we haven't sent yet, discard |
| 3616 | * this segment (RFC793 Section 3.9). |
| 3617 | */ |
| 3618 | if (after(ack, tp->snd_nxt)) |
| 3619 | goto invalid_ack; |
| 3620 | |
Ilpo Järvinen | 2e60529 | 2007-08-02 19:46:58 -0700 | [diff] [blame] | 3621 | if (after(ack, prior_snd_una)) |
| 3622 | flag |= FLAG_SND_UNA_ADVANCED; |
| 3623 | |
Daikichi Osuga | 3fdf3f0 | 2006-08-29 02:01:44 -0700 | [diff] [blame] | 3624 | if (sysctl_tcp_abc) { |
| 3625 | if (icsk->icsk_ca_state < TCP_CA_CWR) |
| 3626 | tp->bytes_acked += ack - prior_snd_una; |
| 3627 | else if (icsk->icsk_ca_state == TCP_CA_Loss) |
| 3628 | /* we assume just one segment left network */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3629 | tp->bytes_acked += min(ack - prior_snd_una, |
| 3630 | tp->mss_cache); |
Daikichi Osuga | 3fdf3f0 | 2006-08-29 02:01:44 -0700 | [diff] [blame] | 3631 | } |
Stephen Hemminger | 9772efb | 2005-11-10 17:09:53 -0800 | [diff] [blame] | 3632 | |
Ilpo Järvinen | c7caf8d | 2007-11-10 21:22:18 -0800 | [diff] [blame] | 3633 | prior_fackets = tp->fackets_out; |
Ilpo Järvinen | 52d3408 | 2007-12-05 02:21:35 -0800 | [diff] [blame] | 3634 | prior_in_flight = tcp_packets_in_flight(tp); |
Ilpo Järvinen | c7caf8d | 2007-11-10 21:22:18 -0800 | [diff] [blame] | 3635 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3636 | if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3637 | /* Window is constant, pure forward advance. |
| 3638 | * No more checks are required. |
| 3639 | * Note, we use the fact that SND.UNA>=SND.WL2. |
| 3640 | */ |
Hantzis Fotis | ee7537b | 2009-03-02 22:42:02 -0800 | [diff] [blame] | 3641 | tcp_update_wl(tp, ack_seq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3642 | tp->snd_una = ack; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3643 | flag |= FLAG_WIN_UPDATE; |
| 3644 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3645 | tcp_ca_event(sk, CA_EVENT_FAST_ACK); |
Stephen Hemminger | 317a76f | 2005-06-23 12:19:55 -0700 | [diff] [blame] | 3646 | |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 3647 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3648 | } else { |
| 3649 | if (ack_seq != TCP_SKB_CB(skb)->end_seq) |
| 3650 | flag |= FLAG_DATA; |
| 3651 | else |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 3652 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3653 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 3654 | flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3655 | |
| 3656 | if (TCP_SKB_CB(skb)->sacked) |
| 3657 | flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); |
| 3658 | |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 3659 | if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb))) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3660 | flag |= FLAG_ECE; |
| 3661 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3662 | tcp_ca_event(sk, CA_EVENT_SLOW_ACK); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3663 | } |
| 3664 | |
| 3665 | /* We passed data and got it acked, remove any soft error |
| 3666 | * log. Something worked... |
| 3667 | */ |
| 3668 | sk->sk_err_soft = 0; |
David S. Miller | 4b53fb6 | 2008-07-23 16:38:45 -0700 | [diff] [blame] | 3669 | icsk->icsk_probes_out = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3670 | tp->rcv_tstamp = tcp_time_stamp; |
| 3671 | prior_packets = tp->packets_out; |
| 3672 | if (!prior_packets) |
| 3673 | goto no_queue; |
| 3674 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3675 | /* See if we can take anything off of the retransmit queue. */ |
Ilpo Järvinen | 33f5f57 | 2008-10-07 14:43:06 -0700 | [diff] [blame] | 3676 | flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3677 | |
Ilpo Järvinen | e1cd8f7 | 2007-11-14 15:55:09 -0800 | [diff] [blame] | 3678 | if (tp->frto_counter) |
| 3679 | frto_cwnd = tcp_process_frto(sk, flag); |
Ilpo Järvinen | 3de9647 | 2007-10-01 15:28:48 -0700 | [diff] [blame] | 3680 | /* Guarantee sacktag reordering detection against wrap-arounds */ |
| 3681 | if (before(tp->frto_highmark, tp->snd_una)) |
| 3682 | tp->frto_highmark = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3683 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3684 | if (tcp_ack_is_dubious(sk, flag)) { |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 3685 | /* Advance CWND, if state allows this. */ |
Ilpo Järvinen | 7c9a4a5 | 2007-02-21 23:08:34 -0800 | [diff] [blame] | 3686 | if ((flag & FLAG_DATA_ACKED) && !frto_cwnd && |
| 3687 | tcp_may_raise_cwnd(sk, flag)) |
Ilpo Järvinen | c3a05c6 | 2007-12-02 00:47:59 +0200 | [diff] [blame] | 3688 | tcp_cong_avoid(sk, ack, prior_in_flight); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3689 | tcp_fastretrans_alert(sk, prior_packets - tp->packets_out, |
| 3690 | flag); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3691 | } else { |
Ilpo Järvinen | 7c9a4a5 | 2007-02-21 23:08:34 -0800 | [diff] [blame] | 3692 | if ((flag & FLAG_DATA_ACKED) && !frto_cwnd) |
Ilpo Järvinen | c3a05c6 | 2007-12-02 00:47:59 +0200 | [diff] [blame] | 3693 | tcp_cong_avoid(sk, ack, prior_in_flight); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3694 | } |
| 3695 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3696 | if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3697 | dst_confirm(sk->sk_dst_cache); |
| 3698 | |
| 3699 | return 1; |
| 3700 | |
| 3701 | no_queue: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3702 | /* If this ack opens up a zero window, clear backoff. It was |
| 3703 | * being used to time the probes, and is probably far higher than |
| 3704 | * it needs to be for normal retransmission. |
| 3705 | */ |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 3706 | if (tcp_send_head(sk)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3707 | tcp_ack_probe(sk); |
| 3708 | return 1; |
| 3709 | |
John Dykstra | 96e0bf4 | 2009-03-22 21:49:57 -0700 | [diff] [blame] | 3710 | invalid_ack: |
| 3711 | SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt); |
| 3712 | return -1; |
| 3713 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3714 | old_ack: |
Ilpo Järvinen | 8aca6cb | 2008-06-04 11:34:22 -0700 | [diff] [blame] | 3715 | if (TCP_SKB_CB(skb)->sacked) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3716 | tcp_sacktag_write_queue(sk, skb, prior_snd_una); |
Ilpo Järvinen | 8aca6cb | 2008-06-04 11:34:22 -0700 | [diff] [blame] | 3717 | if (icsk->icsk_ca_state == TCP_CA_Open) |
| 3718 | tcp_try_keep_open(sk); |
| 3719 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3720 | |
John Dykstra | 96e0bf4 | 2009-03-22 21:49:57 -0700 | [diff] [blame] | 3721 | SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3722 | return 0; |
| 3723 | } |
| 3724 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3725 | /* Look for tcp options. Normally only called on SYN and SYNACK packets. |
| 3726 | * But, this can also be called on packets in the established flow when |
| 3727 | * the fast version below fails. |
| 3728 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3729 | void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 3730 | u8 **hvpp, int estab) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3731 | { |
| 3732 | unsigned char *ptr; |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 3733 | struct tcphdr *th = tcp_hdr(skb); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3734 | int length = (th->doff * 4) - sizeof(struct tcphdr); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3735 | |
| 3736 | ptr = (unsigned char *)(th + 1); |
| 3737 | opt_rx->saw_tstamp = 0; |
| 3738 | |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 3739 | while (length > 0) { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3740 | int opcode = *ptr++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3741 | int opsize; |
| 3742 | |
| 3743 | switch (opcode) { |
Ilpo Järvinen | f038ac8 | 2008-01-03 20:36:55 -0800 | [diff] [blame] | 3744 | case TCPOPT_EOL: |
| 3745 | return; |
| 3746 | case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ |
| 3747 | length--; |
| 3748 | continue; |
| 3749 | default: |
| 3750 | opsize = *ptr++; |
| 3751 | if (opsize < 2) /* "silly options" */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3752 | return; |
Ilpo Järvinen | f038ac8 | 2008-01-03 20:36:55 -0800 | [diff] [blame] | 3753 | if (opsize > length) |
| 3754 | return; /* don't parse partial options */ |
| 3755 | switch (opcode) { |
| 3756 | case TCPOPT_MSS: |
| 3757 | if (opsize == TCPOLEN_MSS && th->syn && !estab) { |
Harvey Harrison | d3e2ce3 | 2008-05-02 16:26:16 -0700 | [diff] [blame] | 3758 | u16 in_mss = get_unaligned_be16(ptr); |
Ilpo Järvinen | f038ac8 | 2008-01-03 20:36:55 -0800 | [diff] [blame] | 3759 | if (in_mss) { |
| 3760 | if (opt_rx->user_mss && |
| 3761 | opt_rx->user_mss < in_mss) |
| 3762 | in_mss = opt_rx->user_mss; |
| 3763 | opt_rx->mss_clamp = in_mss; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3764 | } |
Stephen Hemminger | 3ff50b7 | 2007-04-20 17:09:22 -0700 | [diff] [blame] | 3765 | } |
Ilpo Järvinen | f038ac8 | 2008-01-03 20:36:55 -0800 | [diff] [blame] | 3766 | break; |
| 3767 | case TCPOPT_WINDOW: |
| 3768 | if (opsize == TCPOLEN_WINDOW && th->syn && |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 3769 | !estab && sysctl_tcp_window_scaling) { |
Ilpo Järvinen | f038ac8 | 2008-01-03 20:36:55 -0800 | [diff] [blame] | 3770 | __u8 snd_wscale = *(__u8 *)ptr; |
| 3771 | opt_rx->wscale_ok = 1; |
| 3772 | if (snd_wscale > 14) { |
| 3773 | if (net_ratelimit()) |
| 3774 | printk(KERN_INFO "tcp_parse_options: Illegal window " |
| 3775 | "scaling value %d >14 received.\n", |
| 3776 | snd_wscale); |
| 3777 | snd_wscale = 14; |
| 3778 | } |
| 3779 | opt_rx->snd_wscale = snd_wscale; |
| 3780 | } |
| 3781 | break; |
| 3782 | case TCPOPT_TIMESTAMP: |
| 3783 | if ((opsize == TCPOLEN_TIMESTAMP) && |
| 3784 | ((estab && opt_rx->tstamp_ok) || |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 3785 | (!estab && sysctl_tcp_timestamps))) { |
Ilpo Järvinen | f038ac8 | 2008-01-03 20:36:55 -0800 | [diff] [blame] | 3786 | opt_rx->saw_tstamp = 1; |
Harvey Harrison | d3e2ce3 | 2008-05-02 16:26:16 -0700 | [diff] [blame] | 3787 | opt_rx->rcv_tsval = get_unaligned_be32(ptr); |
| 3788 | opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4); |
Ilpo Järvinen | f038ac8 | 2008-01-03 20:36:55 -0800 | [diff] [blame] | 3789 | } |
| 3790 | break; |
| 3791 | case TCPOPT_SACK_PERM: |
| 3792 | if (opsize == TCPOLEN_SACK_PERM && th->syn && |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 3793 | !estab && sysctl_tcp_sack) { |
Ilpo Järvinen | f038ac8 | 2008-01-03 20:36:55 -0800 | [diff] [blame] | 3794 | opt_rx->sack_ok = 1; |
| 3795 | tcp_sack_reset(opt_rx); |
| 3796 | } |
| 3797 | break; |
Stephen Hemminger | 3ff50b7 | 2007-04-20 17:09:22 -0700 | [diff] [blame] | 3798 | |
Ilpo Järvinen | f038ac8 | 2008-01-03 20:36:55 -0800 | [diff] [blame] | 3799 | case TCPOPT_SACK: |
| 3800 | if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && |
| 3801 | !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && |
| 3802 | opt_rx->sack_ok) { |
| 3803 | TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; |
| 3804 | } |
| 3805 | break; |
| 3806 | #ifdef CONFIG_TCP_MD5SIG |
| 3807 | case TCPOPT_MD5SIG: |
| 3808 | /* |
| 3809 | * The MD5 Hash has already been |
| 3810 | * checked (see tcp_v{4,6}_do_rcv()). |
| 3811 | */ |
| 3812 | break; |
| 3813 | #endif |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 3814 | case TCPOPT_COOKIE: |
| 3815 | /* This option is variable length. |
| 3816 | */ |
| 3817 | switch (opsize) { |
| 3818 | case TCPOLEN_COOKIE_BASE: |
| 3819 | /* not yet implemented */ |
| 3820 | break; |
| 3821 | case TCPOLEN_COOKIE_PAIR: |
| 3822 | /* not yet implemented */ |
| 3823 | break; |
| 3824 | case TCPOLEN_COOKIE_MIN+0: |
| 3825 | case TCPOLEN_COOKIE_MIN+2: |
| 3826 | case TCPOLEN_COOKIE_MIN+4: |
| 3827 | case TCPOLEN_COOKIE_MIN+6: |
| 3828 | case TCPOLEN_COOKIE_MAX: |
| 3829 | /* 16-bit multiple */ |
| 3830 | opt_rx->cookie_plus = opsize; |
| 3831 | *hvpp = ptr; |
| 3832 | default: |
| 3833 | /* ignore option */ |
| 3834 | break; |
| 3835 | }; |
| 3836 | break; |
| 3837 | }; |
Ilpo Järvinen | f038ac8 | 2008-01-03 20:36:55 -0800 | [diff] [blame] | 3838 | |
| 3839 | ptr += opsize-2; |
| 3840 | length -= opsize; |
Stephen Hemminger | 3ff50b7 | 2007-04-20 17:09:22 -0700 | [diff] [blame] | 3841 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3842 | } |
| 3843 | } |
| 3844 | |
Ilpo Järvinen | a4356b2 | 2008-08-23 05:12:29 -0700 | [diff] [blame] | 3845 | static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, struct tcphdr *th) |
| 3846 | { |
| 3847 | __be32 *ptr = (__be32 *)(th + 1); |
| 3848 | |
| 3849 | if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
| 3850 | | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { |
| 3851 | tp->rx_opt.saw_tstamp = 1; |
| 3852 | ++ptr; |
| 3853 | tp->rx_opt.rcv_tsval = ntohl(*ptr); |
| 3854 | ++ptr; |
| 3855 | tp->rx_opt.rcv_tsecr = ntohl(*ptr); |
| 3856 | return 1; |
| 3857 | } |
| 3858 | return 0; |
| 3859 | } |
| 3860 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3861 | /* Fast parse options. This hopes to only see timestamps. |
| 3862 | * If it is wrong it falls back on tcp_parse_options(). |
| 3863 | */ |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 3864 | static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 3865 | struct tcp_sock *tp, u8 **hvpp) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3866 | { |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 3867 | /* In the spirit of fast parsing, compare doff directly to constant |
| 3868 | * values. Because equality is used, short doff can be ignored here. |
| 3869 | */ |
| 3870 | if (th->doff == (sizeof(*th) / 4)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3871 | tp->rx_opt.saw_tstamp = 0; |
| 3872 | return 0; |
| 3873 | } else if (tp->rx_opt.tstamp_ok && |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 3874 | th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) { |
Ilpo Järvinen | a4356b2 | 2008-08-23 05:12:29 -0700 | [diff] [blame] | 3875 | if (tcp_parse_aligned_timestamp(tp, th)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3876 | return 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3877 | } |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 3878 | tcp_parse_options(skb, &tp->rx_opt, hvpp, 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3879 | return 1; |
| 3880 | } |
| 3881 | |
YOSHIFUJI Hideaki | 7d5d552 | 2008-04-17 12:29:53 +0900 | [diff] [blame] | 3882 | #ifdef CONFIG_TCP_MD5SIG |
| 3883 | /* |
| 3884 | * Parse MD5 Signature option |
| 3885 | */ |
| 3886 | u8 *tcp_parse_md5sig_option(struct tcphdr *th) |
| 3887 | { |
| 3888 | int length = (th->doff << 2) - sizeof (*th); |
| 3889 | u8 *ptr = (u8*)(th + 1); |
| 3890 | |
| 3891 | /* If the TCP option is too short, we can short cut */ |
| 3892 | if (length < TCPOLEN_MD5SIG) |
| 3893 | return NULL; |
| 3894 | |
| 3895 | while (length > 0) { |
| 3896 | int opcode = *ptr++; |
| 3897 | int opsize; |
| 3898 | |
| 3899 | switch(opcode) { |
| 3900 | case TCPOPT_EOL: |
| 3901 | return NULL; |
| 3902 | case TCPOPT_NOP: |
| 3903 | length--; |
| 3904 | continue; |
| 3905 | default: |
| 3906 | opsize = *ptr++; |
| 3907 | if (opsize < 2 || opsize > length) |
| 3908 | return NULL; |
| 3909 | if (opcode == TCPOPT_MD5SIG) |
| 3910 | return ptr; |
| 3911 | } |
| 3912 | ptr += opsize - 2; |
| 3913 | length -= opsize; |
| 3914 | } |
| 3915 | return NULL; |
| 3916 | } |
| 3917 | #endif |
| 3918 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3919 | static inline void tcp_store_ts_recent(struct tcp_sock *tp) |
| 3920 | { |
| 3921 | tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; |
James Morris | 9d729f7 | 2007-03-04 16:12:44 -0800 | [diff] [blame] | 3922 | tp->rx_opt.ts_recent_stamp = get_seconds(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3923 | } |
| 3924 | |
| 3925 | static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) |
| 3926 | { |
| 3927 | if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { |
| 3928 | /* PAWS bug workaround wrt. ACK frames, the PAWS discard |
| 3929 | * extra check below makes sure this can only happen |
| 3930 | * for pure ACK frames. -DaveM |
| 3931 | * |
| 3932 | * Not only, also it occurs for expired timestamps. |
| 3933 | */ |
| 3934 | |
Ilpo Järvinen | c887e6d | 2009-03-14 14:23:03 +0000 | [diff] [blame] | 3935 | if (tcp_paws_check(&tp->rx_opt, 0)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3936 | tcp_store_ts_recent(tp); |
| 3937 | } |
| 3938 | } |
| 3939 | |
| 3940 | /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM |
| 3941 | * |
| 3942 | * It is not fatal. If this ACK does _not_ change critical state (seqs, window) |
| 3943 | * it can pass through stack. So, the following predicate verifies that |
| 3944 | * this segment is not used for anything but congestion avoidance or |
| 3945 | * fast retransmit. Moreover, we even are able to eliminate most of such |
| 3946 | * second order effects, if we apply some small "replay" window (~RTO) |
| 3947 | * to timestamp space. |
| 3948 | * |
| 3949 | * All these measures still do not guarantee that we reject wrapped ACKs |
| 3950 | * on networks with high bandwidth, when sequence space is recycled fastly, |
| 3951 | * but it guarantees that such events will be very rare and do not affect |
| 3952 | * connection seriously. This doesn't look nice, but alas, PAWS is really |
| 3953 | * buggy extension. |
| 3954 | * |
| 3955 | * [ Later note. Even worse! It is buggy for segments _with_ data. RFC |
| 3956 | * states that events when retransmit arrives after original data are rare. |
| 3957 | * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is |
| 3958 | * the biggest problem on large power networks even with minor reordering. |
| 3959 | * OK, let's give it small replay window. If peer clock is even 1hz, it is safe |
| 3960 | * up to bandwidth of 18Gigabit/sec. 8) ] |
| 3961 | */ |
| 3962 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3963 | static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3964 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3965 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 3966 | struct tcphdr *th = tcp_hdr(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3967 | u32 seq = TCP_SKB_CB(skb)->seq; |
| 3968 | u32 ack = TCP_SKB_CB(skb)->ack_seq; |
| 3969 | |
| 3970 | return (/* 1. Pure ACK with correct sequence number. */ |
| 3971 | (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && |
| 3972 | |
| 3973 | /* 2. ... and duplicate ACK. */ |
| 3974 | ack == tp->snd_una && |
| 3975 | |
| 3976 | /* 3. ... and does not update window. */ |
| 3977 | !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && |
| 3978 | |
| 3979 | /* 4. ... and sits in replay window. */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3980 | (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3981 | } |
| 3982 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3983 | static inline int tcp_paws_discard(const struct sock *sk, |
| 3984 | const struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3985 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3986 | const struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | c887e6d | 2009-03-14 14:23:03 +0000 | [diff] [blame] | 3987 | |
| 3988 | return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) && |
| 3989 | !tcp_disordered_ack(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3990 | } |
| 3991 | |
| 3992 | /* Check segment sequence number for validity. |
| 3993 | * |
| 3994 | * Segment controls are considered valid, if the segment |
| 3995 | * fits to the window after truncation to the window. Acceptability |
| 3996 | * of data (and SYN, FIN, of course) is checked separately. |
| 3997 | * See tcp_data_queue(), for example. |
| 3998 | * |
| 3999 | * Also, controls (RST is main one) are accepted using RCV.WUP instead |
| 4000 | * of RCV.NXT. Peer still did not advance his SND.UNA when we |
| 4001 | * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. |
| 4002 | * (borrowed from freebsd) |
| 4003 | */ |
| 4004 | |
| 4005 | static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq) |
| 4006 | { |
| 4007 | return !before(end_seq, tp->rcv_wup) && |
| 4008 | !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); |
| 4009 | } |
| 4010 | |
| 4011 | /* When we get a reset we do this. */ |
| 4012 | static void tcp_reset(struct sock *sk) |
| 4013 | { |
| 4014 | /* We want the right error as BSD sees it (and indeed as we do). */ |
| 4015 | switch (sk->sk_state) { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4016 | case TCP_SYN_SENT: |
| 4017 | sk->sk_err = ECONNREFUSED; |
| 4018 | break; |
| 4019 | case TCP_CLOSE_WAIT: |
| 4020 | sk->sk_err = EPIPE; |
| 4021 | break; |
| 4022 | case TCP_CLOSE: |
| 4023 | return; |
| 4024 | default: |
| 4025 | sk->sk_err = ECONNRESET; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4026 | } |
| 4027 | |
| 4028 | if (!sock_flag(sk, SOCK_DEAD)) |
| 4029 | sk->sk_error_report(sk); |
| 4030 | |
| 4031 | tcp_done(sk); |
| 4032 | } |
| 4033 | |
| 4034 | /* |
| 4035 | * Process the FIN bit. This now behaves as it is supposed to work |
| 4036 | * and the FIN takes effect when it is validly part of sequence |
| 4037 | * space. Not before when we get holes. |
| 4038 | * |
| 4039 | * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT |
| 4040 | * (and thence onto LAST-ACK and finally, CLOSE, we never enter |
| 4041 | * TIME-WAIT) |
| 4042 | * |
| 4043 | * If we are in FINWAIT-1, a received FIN indicates simultaneous |
| 4044 | * close and we go into CLOSING (and later onto TIME-WAIT) |
| 4045 | * |
| 4046 | * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. |
| 4047 | */ |
| 4048 | static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th) |
| 4049 | { |
| 4050 | struct tcp_sock *tp = tcp_sk(sk); |
| 4051 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4052 | inet_csk_schedule_ack(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4053 | |
| 4054 | sk->sk_shutdown |= RCV_SHUTDOWN; |
| 4055 | sock_set_flag(sk, SOCK_DONE); |
| 4056 | |
| 4057 | switch (sk->sk_state) { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4058 | case TCP_SYN_RECV: |
| 4059 | case TCP_ESTABLISHED: |
| 4060 | /* Move to CLOSE_WAIT */ |
| 4061 | tcp_set_state(sk, TCP_CLOSE_WAIT); |
| 4062 | inet_csk(sk)->icsk_ack.pingpong = 1; |
| 4063 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4064 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4065 | case TCP_CLOSE_WAIT: |
| 4066 | case TCP_CLOSING: |
| 4067 | /* Received a retransmission of the FIN, do |
| 4068 | * nothing. |
| 4069 | */ |
| 4070 | break; |
| 4071 | case TCP_LAST_ACK: |
| 4072 | /* RFC793: Remain in the LAST-ACK state. */ |
| 4073 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4074 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4075 | case TCP_FIN_WAIT1: |
| 4076 | /* This case occurs when a simultaneous close |
| 4077 | * happens, we must ack the received FIN and |
| 4078 | * enter the CLOSING state. |
| 4079 | */ |
| 4080 | tcp_send_ack(sk); |
| 4081 | tcp_set_state(sk, TCP_CLOSING); |
| 4082 | break; |
| 4083 | case TCP_FIN_WAIT2: |
| 4084 | /* Received a FIN -- send ACK and enter TIME_WAIT. */ |
| 4085 | tcp_send_ack(sk); |
| 4086 | tcp_time_wait(sk, TCP_TIME_WAIT, 0); |
| 4087 | break; |
| 4088 | default: |
| 4089 | /* Only TCP_LISTEN and TCP_CLOSE are left, in these |
| 4090 | * cases we should never reach this piece of code. |
| 4091 | */ |
| 4092 | printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n", |
Harvey Harrison | 0dc4787 | 2008-03-05 20:47:47 -0800 | [diff] [blame] | 4093 | __func__, sk->sk_state); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4094 | break; |
Stephen Hemminger | 3ff50b7 | 2007-04-20 17:09:22 -0700 | [diff] [blame] | 4095 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4096 | |
| 4097 | /* It _is_ possible, that we have something out-of-order _after_ FIN. |
| 4098 | * Probably, we should reset in this case. For now drop them. |
| 4099 | */ |
| 4100 | __skb_queue_purge(&tp->out_of_order_queue); |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 4101 | if (tcp_is_sack(tp)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4102 | tcp_sack_reset(&tp->rx_opt); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 4103 | sk_mem_reclaim(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4104 | |
| 4105 | if (!sock_flag(sk, SOCK_DEAD)) { |
| 4106 | sk->sk_state_change(sk); |
| 4107 | |
| 4108 | /* Do not send POLL_HUP for half duplex close. */ |
| 4109 | if (sk->sk_shutdown == SHUTDOWN_MASK || |
| 4110 | sk->sk_state == TCP_CLOSE) |
Pavel Emelyanov | 8d8ad9d | 2007-11-26 20:10:50 +0800 | [diff] [blame] | 4111 | sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4112 | else |
Pavel Emelyanov | 8d8ad9d | 2007-11-26 20:10:50 +0800 | [diff] [blame] | 4113 | sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4114 | } |
| 4115 | } |
| 4116 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4117 | static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, |
| 4118 | u32 end_seq) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4119 | { |
| 4120 | if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { |
| 4121 | if (before(seq, sp->start_seq)) |
| 4122 | sp->start_seq = seq; |
| 4123 | if (after(end_seq, sp->end_seq)) |
| 4124 | sp->end_seq = end_seq; |
| 4125 | return 1; |
| 4126 | } |
| 4127 | return 0; |
| 4128 | } |
| 4129 | |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 4130 | static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4131 | { |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 4132 | struct tcp_sock *tp = tcp_sk(sk); |
| 4133 | |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 4134 | if (tcp_is_sack(tp) && sysctl_tcp_dsack) { |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 4135 | int mib_idx; |
| 4136 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4137 | if (before(seq, tp->rcv_nxt)) |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 4138 | mib_idx = LINUX_MIB_TCPDSACKOLDSENT; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4139 | else |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 4140 | mib_idx = LINUX_MIB_TCPDSACKOFOSENT; |
| 4141 | |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 4142 | NET_INC_STATS_BH(sock_net(sk), mib_idx); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4143 | |
| 4144 | tp->rx_opt.dsack = 1; |
| 4145 | tp->duplicate_sack[0].start_seq = seq; |
| 4146 | tp->duplicate_sack[0].end_seq = end_seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4147 | } |
| 4148 | } |
| 4149 | |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 4150 | static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4151 | { |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 4152 | struct tcp_sock *tp = tcp_sk(sk); |
| 4153 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4154 | if (!tp->rx_opt.dsack) |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 4155 | tcp_dsack_set(sk, seq, end_seq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4156 | else |
| 4157 | tcp_sack_extend(tp->duplicate_sack, seq, end_seq); |
| 4158 | } |
| 4159 | |
| 4160 | static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb) |
| 4161 | { |
| 4162 | struct tcp_sock *tp = tcp_sk(sk); |
| 4163 | |
| 4164 | if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && |
| 4165 | before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 4166 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4167 | tcp_enter_quickack_mode(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4168 | |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 4169 | if (tcp_is_sack(tp) && sysctl_tcp_dsack) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4170 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; |
| 4171 | |
| 4172 | if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) |
| 4173 | end_seq = tp->rcv_nxt; |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 4174 | tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4175 | } |
| 4176 | } |
| 4177 | |
| 4178 | tcp_send_ack(sk); |
| 4179 | } |
| 4180 | |
| 4181 | /* These routines update the SACK block as out-of-order packets arrive or |
| 4182 | * in-order packets close up the sequence space. |
| 4183 | */ |
| 4184 | static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) |
| 4185 | { |
| 4186 | int this_sack; |
| 4187 | struct tcp_sack_block *sp = &tp->selective_acks[0]; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4188 | struct tcp_sack_block *swalk = sp + 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4189 | |
| 4190 | /* See if the recent change to the first SACK eats into |
| 4191 | * or hits the sequence space of other SACK blocks, if so coalesce. |
| 4192 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4193 | for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4194 | if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { |
| 4195 | int i; |
| 4196 | |
| 4197 | /* Zap SWALK, by moving every further SACK up by one slot. |
| 4198 | * Decrease num_sacks. |
| 4199 | */ |
| 4200 | tp->rx_opt.num_sacks--; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4201 | for (i = this_sack; i < tp->rx_opt.num_sacks; i++) |
| 4202 | sp[i] = sp[i + 1]; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4203 | continue; |
| 4204 | } |
| 4205 | this_sack++, swalk++; |
| 4206 | } |
| 4207 | } |
| 4208 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4209 | static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) |
| 4210 | { |
| 4211 | struct tcp_sock *tp = tcp_sk(sk); |
| 4212 | struct tcp_sack_block *sp = &tp->selective_acks[0]; |
| 4213 | int cur_sacks = tp->rx_opt.num_sacks; |
| 4214 | int this_sack; |
| 4215 | |
| 4216 | if (!cur_sacks) |
| 4217 | goto new_sack; |
| 4218 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4219 | for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4220 | if (tcp_sack_extend(sp, seq, end_seq)) { |
| 4221 | /* Rotate this_sack to the first one. */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4222 | for (; this_sack > 0; this_sack--, sp--) |
Ilpo Järvinen | a0bffff | 2009-03-21 13:36:17 -0700 | [diff] [blame] | 4223 | swap(*sp, *(sp - 1)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4224 | if (cur_sacks > 1) |
| 4225 | tcp_sack_maybe_coalesce(tp); |
| 4226 | return; |
| 4227 | } |
| 4228 | } |
| 4229 | |
| 4230 | /* Could not find an adjacent existing SACK, build a new one, |
| 4231 | * put it at the front, and shift everyone else down. We |
| 4232 | * always know there is at least one SACK present already here. |
| 4233 | * |
| 4234 | * If the sack array is full, forget about the last one. |
| 4235 | */ |
Adam Langley | 4389dde | 2008-07-19 00:07:02 -0700 | [diff] [blame] | 4236 | if (this_sack >= TCP_NUM_SACKS) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4237 | this_sack--; |
| 4238 | tp->rx_opt.num_sacks--; |
| 4239 | sp--; |
| 4240 | } |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 4241 | for (; this_sack > 0; this_sack--, sp--) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4242 | *sp = *(sp - 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4243 | |
| 4244 | new_sack: |
| 4245 | /* Build the new head SACK, and we're done. */ |
| 4246 | sp->start_seq = seq; |
| 4247 | sp->end_seq = end_seq; |
| 4248 | tp->rx_opt.num_sacks++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4249 | } |
| 4250 | |
| 4251 | /* RCV.NXT advances, some SACKs should be eaten. */ |
| 4252 | |
| 4253 | static void tcp_sack_remove(struct tcp_sock *tp) |
| 4254 | { |
| 4255 | struct tcp_sack_block *sp = &tp->selective_acks[0]; |
| 4256 | int num_sacks = tp->rx_opt.num_sacks; |
| 4257 | int this_sack; |
| 4258 | |
| 4259 | /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ |
David S. Miller | b03efcf | 2005-07-08 14:57:23 -0700 | [diff] [blame] | 4260 | if (skb_queue_empty(&tp->out_of_order_queue)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4261 | tp->rx_opt.num_sacks = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4262 | return; |
| 4263 | } |
| 4264 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4265 | for (this_sack = 0; this_sack < num_sacks;) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4266 | /* Check if the start of the sack is covered by RCV.NXT. */ |
| 4267 | if (!before(tp->rcv_nxt, sp->start_seq)) { |
| 4268 | int i; |
| 4269 | |
| 4270 | /* RCV.NXT must cover all the block! */ |
Ilpo Järvinen | 547b792 | 2008-07-25 21:43:18 -0700 | [diff] [blame] | 4271 | WARN_ON(before(tp->rcv_nxt, sp->end_seq)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4272 | |
| 4273 | /* Zap this SACK, by moving forward any other SACKS. */ |
| 4274 | for (i=this_sack+1; i < num_sacks; i++) |
| 4275 | tp->selective_acks[i-1] = tp->selective_acks[i]; |
| 4276 | num_sacks--; |
| 4277 | continue; |
| 4278 | } |
| 4279 | this_sack++; |
| 4280 | sp++; |
| 4281 | } |
Ilpo Järvinen | 5861f8e | 2009-03-14 14:23:01 +0000 | [diff] [blame] | 4282 | tp->rx_opt.num_sacks = num_sacks; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4283 | } |
| 4284 | |
| 4285 | /* This one checks to see if we can put data from the |
| 4286 | * out_of_order queue into the receive_queue. |
| 4287 | */ |
| 4288 | static void tcp_ofo_queue(struct sock *sk) |
| 4289 | { |
| 4290 | struct tcp_sock *tp = tcp_sk(sk); |
| 4291 | __u32 dsack_high = tp->rcv_nxt; |
| 4292 | struct sk_buff *skb; |
| 4293 | |
| 4294 | while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { |
| 4295 | if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) |
| 4296 | break; |
| 4297 | |
| 4298 | if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { |
| 4299 | __u32 dsack = dsack_high; |
| 4300 | if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) |
| 4301 | dsack_high = TCP_SKB_CB(skb)->end_seq; |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 4302 | tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4303 | } |
| 4304 | |
| 4305 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { |
| 4306 | SOCK_DEBUG(sk, "ofo packet was already received \n"); |
David S. Miller | 8728b83 | 2005-08-09 19:25:21 -0700 | [diff] [blame] | 4307 | __skb_unlink(skb, &tp->out_of_order_queue); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4308 | __kfree_skb(skb); |
| 4309 | continue; |
| 4310 | } |
| 4311 | SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", |
| 4312 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, |
| 4313 | TCP_SKB_CB(skb)->end_seq); |
| 4314 | |
David S. Miller | 8728b83 | 2005-08-09 19:25:21 -0700 | [diff] [blame] | 4315 | __skb_unlink(skb, &tp->out_of_order_queue); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4316 | __skb_queue_tail(&sk->sk_receive_queue, skb); |
| 4317 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 4318 | if (tcp_hdr(skb)->fin) |
| 4319 | tcp_fin(skb, sk, tcp_hdr(skb)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4320 | } |
| 4321 | } |
| 4322 | |
Vitaliy Gusev | 56f367b | 2008-04-15 20:26:34 -0700 | [diff] [blame] | 4323 | static int tcp_prune_ofo_queue(struct sock *sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4324 | static int tcp_prune_queue(struct sock *sk); |
| 4325 | |
Vitaliy Gusev | b000cd3 | 2008-04-15 00:33:38 -0700 | [diff] [blame] | 4326 | static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size) |
| 4327 | { |
| 4328 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || |
| 4329 | !sk_rmem_schedule(sk, size)) { |
| 4330 | |
| 4331 | if (tcp_prune_queue(sk) < 0) |
| 4332 | return -1; |
| 4333 | |
| 4334 | if (!sk_rmem_schedule(sk, size)) { |
Vitaliy Gusev | 56f367b | 2008-04-15 20:26:34 -0700 | [diff] [blame] | 4335 | if (!tcp_prune_ofo_queue(sk)) |
| 4336 | return -1; |
| 4337 | |
Vitaliy Gusev | b000cd3 | 2008-04-15 00:33:38 -0700 | [diff] [blame] | 4338 | if (!sk_rmem_schedule(sk, size)) |
| 4339 | return -1; |
| 4340 | } |
| 4341 | } |
| 4342 | return 0; |
| 4343 | } |
| 4344 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4345 | static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) |
| 4346 | { |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 4347 | struct tcphdr *th = tcp_hdr(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4348 | struct tcp_sock *tp = tcp_sk(sk); |
| 4349 | int eaten = -1; |
| 4350 | |
| 4351 | if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) |
| 4352 | goto drop; |
| 4353 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4354 | __skb_pull(skb, th->doff * 4); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4355 | |
| 4356 | TCP_ECN_accept_cwr(tp, skb); |
| 4357 | |
Ilpo Järvinen | 5861f8e | 2009-03-14 14:23:01 +0000 | [diff] [blame] | 4358 | tp->rx_opt.dsack = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4359 | |
| 4360 | /* Queue data for delivery to the user. |
| 4361 | * Packets in sequence go to the receive queue. |
| 4362 | * Out of sequence packets to the out_of_order_queue. |
| 4363 | */ |
| 4364 | if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { |
| 4365 | if (tcp_receive_window(tp) == 0) |
| 4366 | goto out_of_window; |
| 4367 | |
| 4368 | /* Ok. In sequence. In window. */ |
| 4369 | if (tp->ucopy.task == current && |
| 4370 | tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && |
| 4371 | sock_owned_by_user(sk) && !tp->urg_data) { |
| 4372 | int chunk = min_t(unsigned int, skb->len, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4373 | tp->ucopy.len); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4374 | |
| 4375 | __set_current_state(TASK_RUNNING); |
| 4376 | |
| 4377 | local_bh_enable(); |
| 4378 | if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { |
| 4379 | tp->ucopy.len -= chunk; |
| 4380 | tp->copied_seq += chunk; |
| 4381 | eaten = (chunk == skb->len && !th->fin); |
| 4382 | tcp_rcv_space_adjust(sk); |
| 4383 | } |
| 4384 | local_bh_disable(); |
| 4385 | } |
| 4386 | |
| 4387 | if (eaten <= 0) { |
| 4388 | queue_and_out: |
| 4389 | if (eaten < 0 && |
Vitaliy Gusev | b000cd3 | 2008-04-15 00:33:38 -0700 | [diff] [blame] | 4390 | tcp_try_rmem_schedule(sk, skb->truesize)) |
| 4391 | goto drop; |
| 4392 | |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 4393 | skb_set_owner_r(skb, sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4394 | __skb_queue_tail(&sk->sk_receive_queue, skb); |
| 4395 | } |
| 4396 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 4397 | if (skb->len) |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 4398 | tcp_event_data_recv(sk, skb); |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 4399 | if (th->fin) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4400 | tcp_fin(skb, sk, th); |
| 4401 | |
David S. Miller | b03efcf | 2005-07-08 14:57:23 -0700 | [diff] [blame] | 4402 | if (!skb_queue_empty(&tp->out_of_order_queue)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4403 | tcp_ofo_queue(sk); |
| 4404 | |
| 4405 | /* RFC2581. 4.2. SHOULD send immediate ACK, when |
| 4406 | * gap in queue is filled. |
| 4407 | */ |
David S. Miller | b03efcf | 2005-07-08 14:57:23 -0700 | [diff] [blame] | 4408 | if (skb_queue_empty(&tp->out_of_order_queue)) |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4409 | inet_csk(sk)->icsk_ack.pingpong = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4410 | } |
| 4411 | |
| 4412 | if (tp->rx_opt.num_sacks) |
| 4413 | tcp_sack_remove(tp); |
| 4414 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 4415 | tcp_fast_path_check(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4416 | |
| 4417 | if (eaten > 0) |
| 4418 | __kfree_skb(skb); |
| 4419 | else if (!sock_flag(sk, SOCK_DEAD)) |
| 4420 | sk->sk_data_ready(sk, 0); |
| 4421 | return; |
| 4422 | } |
| 4423 | |
| 4424 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { |
| 4425 | /* A retransmit, 2nd most common case. Force an immediate ack. */ |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 4426 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 4427 | tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4428 | |
| 4429 | out_of_window: |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4430 | tcp_enter_quickack_mode(sk); |
| 4431 | inet_csk_schedule_ack(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4432 | drop: |
| 4433 | __kfree_skb(skb); |
| 4434 | return; |
| 4435 | } |
| 4436 | |
| 4437 | /* Out of window. F.e. zero window probe. */ |
| 4438 | if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) |
| 4439 | goto out_of_window; |
| 4440 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4441 | tcp_enter_quickack_mode(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4442 | |
| 4443 | if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { |
| 4444 | /* Partial packet, seq < rcv_next < end_seq */ |
| 4445 | SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", |
| 4446 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, |
| 4447 | TCP_SKB_CB(skb)->end_seq); |
| 4448 | |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 4449 | tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4450 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4451 | /* If window is closed, drop tail of packet. But after |
| 4452 | * remembering D-SACK for its head made in previous line. |
| 4453 | */ |
| 4454 | if (!tcp_receive_window(tp)) |
| 4455 | goto out_of_window; |
| 4456 | goto queue_and_out; |
| 4457 | } |
| 4458 | |
| 4459 | TCP_ECN_check_ce(tp, skb); |
| 4460 | |
Vitaliy Gusev | b000cd3 | 2008-04-15 00:33:38 -0700 | [diff] [blame] | 4461 | if (tcp_try_rmem_schedule(sk, skb->truesize)) |
| 4462 | goto drop; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4463 | |
| 4464 | /* Disable header prediction. */ |
| 4465 | tp->pred_flags = 0; |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4466 | inet_csk_schedule_ack(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4467 | |
| 4468 | SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", |
| 4469 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); |
| 4470 | |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 4471 | skb_set_owner_r(skb, sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4472 | |
| 4473 | if (!skb_peek(&tp->out_of_order_queue)) { |
| 4474 | /* Initial out of order segment, build 1 SACK. */ |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 4475 | if (tcp_is_sack(tp)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4476 | tp->rx_opt.num_sacks = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4477 | tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; |
| 4478 | tp->selective_acks[0].end_seq = |
| 4479 | TCP_SKB_CB(skb)->end_seq; |
| 4480 | } |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4481 | __skb_queue_head(&tp->out_of_order_queue, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4482 | } else { |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4483 | struct sk_buff *skb1 = skb_peek_tail(&tp->out_of_order_queue); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4484 | u32 seq = TCP_SKB_CB(skb)->seq; |
| 4485 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; |
| 4486 | |
| 4487 | if (seq == TCP_SKB_CB(skb1)->end_seq) { |
Gerrit Renker | 7de6c03 | 2008-04-14 00:05:09 -0700 | [diff] [blame] | 4488 | __skb_queue_after(&tp->out_of_order_queue, skb1, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4489 | |
| 4490 | if (!tp->rx_opt.num_sacks || |
| 4491 | tp->selective_acks[0].end_seq != seq) |
| 4492 | goto add_sack; |
| 4493 | |
| 4494 | /* Common case: data arrive in order after hole. */ |
| 4495 | tp->selective_acks[0].end_seq = end_seq; |
| 4496 | return; |
| 4497 | } |
| 4498 | |
| 4499 | /* Find place to insert this segment. */ |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4500 | while (1) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4501 | if (!after(TCP_SKB_CB(skb1)->seq, seq)) |
| 4502 | break; |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4503 | if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) { |
| 4504 | skb1 = NULL; |
| 4505 | break; |
| 4506 | } |
| 4507 | skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1); |
| 4508 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4509 | |
| 4510 | /* Do skb overlap to previous one? */ |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4511 | if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4512 | if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { |
| 4513 | /* All the bits are present. Drop. */ |
| 4514 | __kfree_skb(skb); |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 4515 | tcp_dsack_set(sk, seq, end_seq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4516 | goto add_sack; |
| 4517 | } |
| 4518 | if (after(seq, TCP_SKB_CB(skb1)->seq)) { |
| 4519 | /* Partial overlap. */ |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 4520 | tcp_dsack_set(sk, seq, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4521 | TCP_SKB_CB(skb1)->end_seq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4522 | } else { |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4523 | if (skb_queue_is_first(&tp->out_of_order_queue, |
| 4524 | skb1)) |
| 4525 | skb1 = NULL; |
| 4526 | else |
| 4527 | skb1 = skb_queue_prev( |
| 4528 | &tp->out_of_order_queue, |
| 4529 | skb1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4530 | } |
| 4531 | } |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4532 | if (!skb1) |
| 4533 | __skb_queue_head(&tp->out_of_order_queue, skb); |
| 4534 | else |
| 4535 | __skb_queue_after(&tp->out_of_order_queue, skb1, skb); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4536 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4537 | /* And clean segments covered by new one as whole. */ |
Ilpo Järvinen | 2df9001 | 2009-05-29 15:02:29 -0700 | [diff] [blame] | 4538 | while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) { |
| 4539 | skb1 = skb_queue_next(&tp->out_of_order_queue, skb); |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4540 | |
Ilpo Järvinen | 2df9001 | 2009-05-29 15:02:29 -0700 | [diff] [blame] | 4541 | if (!after(end_seq, TCP_SKB_CB(skb1)->seq)) |
| 4542 | break; |
| 4543 | if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 4544 | tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, |
Ilpo Järvinen | 2df9001 | 2009-05-29 15:02:29 -0700 | [diff] [blame] | 4545 | end_seq); |
| 4546 | break; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4547 | } |
Ilpo Järvinen | 2df9001 | 2009-05-29 15:02:29 -0700 | [diff] [blame] | 4548 | __skb_unlink(skb1, &tp->out_of_order_queue); |
| 4549 | tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, |
| 4550 | TCP_SKB_CB(skb1)->end_seq); |
| 4551 | __kfree_skb(skb1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4552 | } |
| 4553 | |
| 4554 | add_sack: |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 4555 | if (tcp_is_sack(tp)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4556 | tcp_sack_new_ofo_skb(sk, seq, end_seq); |
| 4557 | } |
| 4558 | } |
| 4559 | |
Ilpo Järvinen | 2cf4663 | 2008-08-23 05:11:41 -0700 | [diff] [blame] | 4560 | static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb, |
| 4561 | struct sk_buff_head *list) |
| 4562 | { |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4563 | struct sk_buff *next = NULL; |
| 4564 | |
| 4565 | if (!skb_queue_is_last(list, skb)) |
| 4566 | next = skb_queue_next(list, skb); |
Ilpo Järvinen | 2cf4663 | 2008-08-23 05:11:41 -0700 | [diff] [blame] | 4567 | |
| 4568 | __skb_unlink(skb, list); |
| 4569 | __kfree_skb(skb); |
| 4570 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED); |
| 4571 | |
| 4572 | return next; |
| 4573 | } |
| 4574 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4575 | /* Collapse contiguous sequence of skbs head..tail with |
| 4576 | * sequence numbers start..end. |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4577 | * |
| 4578 | * If tail is NULL, this means until the end of the list. |
| 4579 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4580 | * Segments with FIN/SYN are not collapsed (only because this |
| 4581 | * simplifies code) |
| 4582 | */ |
| 4583 | static void |
David S. Miller | 8728b83 | 2005-08-09 19:25:21 -0700 | [diff] [blame] | 4584 | tcp_collapse(struct sock *sk, struct sk_buff_head *list, |
| 4585 | struct sk_buff *head, struct sk_buff *tail, |
| 4586 | u32 start, u32 end) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4587 | { |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4588 | struct sk_buff *skb, *n; |
| 4589 | bool end_of_skbs; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4590 | |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 4591 | /* First, check that queue is collapsible and find |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4592 | * the point where collapsing can be useful. */ |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4593 | skb = head; |
| 4594 | restart: |
| 4595 | end_of_skbs = true; |
| 4596 | skb_queue_walk_from_safe(list, skb, n) { |
| 4597 | if (skb == tail) |
| 4598 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4599 | /* No new bits? It is possible on ofo queue. */ |
| 4600 | if (!before(start, TCP_SKB_CB(skb)->end_seq)) { |
Ilpo Järvinen | 2cf4663 | 2008-08-23 05:11:41 -0700 | [diff] [blame] | 4601 | skb = tcp_collapse_one(sk, skb, list); |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4602 | if (!skb) |
| 4603 | break; |
| 4604 | goto restart; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4605 | } |
| 4606 | |
| 4607 | /* The first skb to collapse is: |
| 4608 | * - not SYN/FIN and |
| 4609 | * - bloated or contains data before "start" or |
| 4610 | * overlaps to the next one. |
| 4611 | */ |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 4612 | if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin && |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4613 | (tcp_win_from_space(skb->truesize) > skb->len || |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4614 | before(TCP_SKB_CB(skb)->seq, start))) { |
| 4615 | end_of_skbs = false; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4616 | break; |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4617 | } |
| 4618 | |
| 4619 | if (!skb_queue_is_last(list, skb)) { |
| 4620 | struct sk_buff *next = skb_queue_next(list, skb); |
| 4621 | if (next != tail && |
| 4622 | TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) { |
| 4623 | end_of_skbs = false; |
| 4624 | break; |
| 4625 | } |
| 4626 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4627 | |
| 4628 | /* Decided to skip this, advance start seq. */ |
| 4629 | start = TCP_SKB_CB(skb)->end_seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4630 | } |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4631 | if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4632 | return; |
| 4633 | |
| 4634 | while (before(start, end)) { |
| 4635 | struct sk_buff *nskb; |
Chuck Lever | c2636b4 | 2007-10-23 21:07:32 -0700 | [diff] [blame] | 4636 | unsigned int header = skb_headroom(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4637 | int copy = SKB_MAX_ORDER(header, 0); |
| 4638 | |
| 4639 | /* Too big header? This can happen with IPv6. */ |
| 4640 | if (copy < 0) |
| 4641 | return; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4642 | if (end - start < copy) |
| 4643 | copy = end - start; |
| 4644 | nskb = alloc_skb(copy + header, GFP_ATOMIC); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4645 | if (!nskb) |
| 4646 | return; |
Arnaldo Carvalho de Melo | c51957d | 2007-03-10 12:47:22 -0300 | [diff] [blame] | 4647 | |
Arnaldo Carvalho de Melo | 98e399f | 2007-03-19 15:33:04 -0700 | [diff] [blame] | 4648 | skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head); |
Arnaldo Carvalho de Melo | 9c70220 | 2007-04-25 18:04:18 -0700 | [diff] [blame] | 4649 | skb_set_network_header(nskb, (skb_network_header(skb) - |
| 4650 | skb->head)); |
| 4651 | skb_set_transport_header(nskb, (skb_transport_header(skb) - |
| 4652 | skb->head)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4653 | skb_reserve(nskb, header); |
| 4654 | memcpy(nskb->head, skb->head, header); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4655 | memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); |
| 4656 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; |
David S. Miller | 43f59c8 | 2008-09-21 21:28:51 -0700 | [diff] [blame] | 4657 | __skb_queue_before(list, skb, nskb); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 4658 | skb_set_owner_r(nskb, sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4659 | |
| 4660 | /* Copy data, releasing collapsed skbs. */ |
| 4661 | while (copy > 0) { |
| 4662 | int offset = start - TCP_SKB_CB(skb)->seq; |
| 4663 | int size = TCP_SKB_CB(skb)->end_seq - start; |
| 4664 | |
Kris Katterjohn | 09a6266 | 2006-01-08 22:24:28 -0800 | [diff] [blame] | 4665 | BUG_ON(offset < 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4666 | if (size > 0) { |
| 4667 | size = min(copy, size); |
| 4668 | if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) |
| 4669 | BUG(); |
| 4670 | TCP_SKB_CB(nskb)->end_seq += size; |
| 4671 | copy -= size; |
| 4672 | start += size; |
| 4673 | } |
| 4674 | if (!before(start, TCP_SKB_CB(skb)->end_seq)) { |
Ilpo Järvinen | 2cf4663 | 2008-08-23 05:11:41 -0700 | [diff] [blame] | 4675 | skb = tcp_collapse_one(sk, skb, list); |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4676 | if (!skb || |
| 4677 | skb == tail || |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 4678 | tcp_hdr(skb)->syn || |
| 4679 | tcp_hdr(skb)->fin) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4680 | return; |
| 4681 | } |
| 4682 | } |
| 4683 | } |
| 4684 | } |
| 4685 | |
| 4686 | /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs |
| 4687 | * and tcp_collapse() them until all the queue is collapsed. |
| 4688 | */ |
| 4689 | static void tcp_collapse_ofo_queue(struct sock *sk) |
| 4690 | { |
| 4691 | struct tcp_sock *tp = tcp_sk(sk); |
| 4692 | struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); |
| 4693 | struct sk_buff *head; |
| 4694 | u32 start, end; |
| 4695 | |
| 4696 | if (skb == NULL) |
| 4697 | return; |
| 4698 | |
| 4699 | start = TCP_SKB_CB(skb)->seq; |
| 4700 | end = TCP_SKB_CB(skb)->end_seq; |
| 4701 | head = skb; |
| 4702 | |
| 4703 | for (;;) { |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4704 | struct sk_buff *next = NULL; |
| 4705 | |
| 4706 | if (!skb_queue_is_last(&tp->out_of_order_queue, skb)) |
| 4707 | next = skb_queue_next(&tp->out_of_order_queue, skb); |
| 4708 | skb = next; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4709 | |
| 4710 | /* Segment is terminated when we see gap or when |
| 4711 | * we are at the end of all the queue. */ |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4712 | if (!skb || |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4713 | after(TCP_SKB_CB(skb)->seq, end) || |
| 4714 | before(TCP_SKB_CB(skb)->end_seq, start)) { |
David S. Miller | 8728b83 | 2005-08-09 19:25:21 -0700 | [diff] [blame] | 4715 | tcp_collapse(sk, &tp->out_of_order_queue, |
| 4716 | head, skb, start, end); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4717 | head = skb; |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4718 | if (!skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4719 | break; |
| 4720 | /* Start new segment */ |
| 4721 | start = TCP_SKB_CB(skb)->seq; |
| 4722 | end = TCP_SKB_CB(skb)->end_seq; |
| 4723 | } else { |
| 4724 | if (before(TCP_SKB_CB(skb)->seq, start)) |
| 4725 | start = TCP_SKB_CB(skb)->seq; |
| 4726 | if (after(TCP_SKB_CB(skb)->end_seq, end)) |
| 4727 | end = TCP_SKB_CB(skb)->end_seq; |
| 4728 | } |
| 4729 | } |
| 4730 | } |
| 4731 | |
Vitaliy Gusev | b000cd3 | 2008-04-15 00:33:38 -0700 | [diff] [blame] | 4732 | /* |
| 4733 | * Purge the out-of-order queue. |
Vitaliy Gusev | 56f367b | 2008-04-15 20:26:34 -0700 | [diff] [blame] | 4734 | * Return true if queue was pruned. |
Vitaliy Gusev | b000cd3 | 2008-04-15 00:33:38 -0700 | [diff] [blame] | 4735 | */ |
Vitaliy Gusev | 56f367b | 2008-04-15 20:26:34 -0700 | [diff] [blame] | 4736 | static int tcp_prune_ofo_queue(struct sock *sk) |
Vitaliy Gusev | b000cd3 | 2008-04-15 00:33:38 -0700 | [diff] [blame] | 4737 | { |
| 4738 | struct tcp_sock *tp = tcp_sk(sk); |
Vitaliy Gusev | 56f367b | 2008-04-15 20:26:34 -0700 | [diff] [blame] | 4739 | int res = 0; |
Vitaliy Gusev | b000cd3 | 2008-04-15 00:33:38 -0700 | [diff] [blame] | 4740 | |
| 4741 | if (!skb_queue_empty(&tp->out_of_order_queue)) { |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 4742 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED); |
Vitaliy Gusev | b000cd3 | 2008-04-15 00:33:38 -0700 | [diff] [blame] | 4743 | __skb_queue_purge(&tp->out_of_order_queue); |
| 4744 | |
| 4745 | /* Reset SACK state. A conforming SACK implementation will |
| 4746 | * do the same at a timeout based retransmit. When a connection |
| 4747 | * is in a sad state like this, we care only about integrity |
| 4748 | * of the connection not performance. |
| 4749 | */ |
| 4750 | if (tp->rx_opt.sack_ok) |
| 4751 | tcp_sack_reset(&tp->rx_opt); |
| 4752 | sk_mem_reclaim(sk); |
Vitaliy Gusev | 56f367b | 2008-04-15 20:26:34 -0700 | [diff] [blame] | 4753 | res = 1; |
Vitaliy Gusev | b000cd3 | 2008-04-15 00:33:38 -0700 | [diff] [blame] | 4754 | } |
Vitaliy Gusev | 56f367b | 2008-04-15 20:26:34 -0700 | [diff] [blame] | 4755 | return res; |
Vitaliy Gusev | b000cd3 | 2008-04-15 00:33:38 -0700 | [diff] [blame] | 4756 | } |
| 4757 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4758 | /* Reduce allocated memory if we can, trying to get |
| 4759 | * the socket within its memory limits again. |
| 4760 | * |
| 4761 | * Return less than zero if we should start dropping frames |
| 4762 | * until the socket owning process reads some of the data |
| 4763 | * to stabilize the situation. |
| 4764 | */ |
| 4765 | static int tcp_prune_queue(struct sock *sk) |
| 4766 | { |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4767 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4768 | |
| 4769 | SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); |
| 4770 | |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 4771 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4772 | |
| 4773 | if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 4774 | tcp_clamp_window(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4775 | else if (tcp_memory_pressure) |
| 4776 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); |
| 4777 | |
| 4778 | tcp_collapse_ofo_queue(sk); |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4779 | if (!skb_queue_empty(&sk->sk_receive_queue)) |
| 4780 | tcp_collapse(sk, &sk->sk_receive_queue, |
| 4781 | skb_peek(&sk->sk_receive_queue), |
| 4782 | NULL, |
| 4783 | tp->copied_seq, tp->rcv_nxt); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 4784 | sk_mem_reclaim(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4785 | |
| 4786 | if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) |
| 4787 | return 0; |
| 4788 | |
| 4789 | /* Collapsing did not help, destructive actions follow. |
| 4790 | * This must not ever occur. */ |
| 4791 | |
Vitaliy Gusev | b000cd3 | 2008-04-15 00:33:38 -0700 | [diff] [blame] | 4792 | tcp_prune_ofo_queue(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4793 | |
| 4794 | if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) |
| 4795 | return 0; |
| 4796 | |
| 4797 | /* If we are really being abused, tell the caller to silently |
| 4798 | * drop receive data on the floor. It will get retransmitted |
| 4799 | * and hopefully then we'll have sufficient space. |
| 4800 | */ |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 4801 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4802 | |
| 4803 | /* Massive buffer overcommit. */ |
| 4804 | tp->pred_flags = 0; |
| 4805 | return -1; |
| 4806 | } |
| 4807 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4808 | /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. |
| 4809 | * As additional protections, we do not touch cwnd in retransmission phases, |
| 4810 | * and if application hit its sndbuf limit recently. |
| 4811 | */ |
| 4812 | void tcp_cwnd_application_limited(struct sock *sk) |
| 4813 | { |
| 4814 | struct tcp_sock *tp = tcp_sk(sk); |
| 4815 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 4816 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4817 | sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { |
| 4818 | /* Limited by application or receiver window. */ |
Ilpo Järvinen | d254bcd | 2006-08-04 16:57:42 -0700 | [diff] [blame] | 4819 | u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); |
| 4820 | u32 win_used = max(tp->snd_cwnd_used, init_win); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4821 | if (win_used < tp->snd_cwnd) { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 4822 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4823 | tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; |
| 4824 | } |
| 4825 | tp->snd_cwnd_used = 0; |
| 4826 | } |
| 4827 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 4828 | } |
| 4829 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 4830 | static int tcp_should_expand_sndbuf(struct sock *sk) |
David S. Miller | 0d9901d | 2005-07-05 15:21:10 -0700 | [diff] [blame] | 4831 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 4832 | struct tcp_sock *tp = tcp_sk(sk); |
| 4833 | |
David S. Miller | 0d9901d | 2005-07-05 15:21:10 -0700 | [diff] [blame] | 4834 | /* If the user specified a specific send buffer setting, do |
| 4835 | * not modify it. |
| 4836 | */ |
| 4837 | if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) |
| 4838 | return 0; |
| 4839 | |
| 4840 | /* If we are under global TCP memory pressure, do not expand. */ |
| 4841 | if (tcp_memory_pressure) |
| 4842 | return 0; |
| 4843 | |
| 4844 | /* If we are under soft global TCP memory pressure, do not expand. */ |
| 4845 | if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0]) |
| 4846 | return 0; |
| 4847 | |
| 4848 | /* If we filled the congestion window, do not expand. */ |
| 4849 | if (tp->packets_out >= tp->snd_cwnd) |
| 4850 | return 0; |
| 4851 | |
| 4852 | return 1; |
| 4853 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4854 | |
| 4855 | /* When incoming ACK allowed to free some skb from write_queue, |
| 4856 | * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket |
| 4857 | * on the exit from tcp input handler. |
| 4858 | * |
| 4859 | * PROBLEM: sndbuf expansion does not work well with largesend. |
| 4860 | */ |
| 4861 | static void tcp_new_space(struct sock *sk) |
| 4862 | { |
| 4863 | struct tcp_sock *tp = tcp_sk(sk); |
| 4864 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 4865 | if (tcp_should_expand_sndbuf(sk)) { |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4866 | int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + |
Ilpo Järvinen | 4a7e560 | 2008-10-07 14:43:31 -0700 | [diff] [blame] | 4867 | MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); |
| 4868 | int demanded = max_t(unsigned int, tp->snd_cwnd, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4869 | tp->reordering + 1); |
| 4870 | sndmem *= 2 * demanded; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4871 | if (sndmem > sk->sk_sndbuf) |
| 4872 | sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); |
| 4873 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 4874 | } |
| 4875 | |
| 4876 | sk->sk_write_space(sk); |
| 4877 | } |
| 4878 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 4879 | static void tcp_check_space(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4880 | { |
| 4881 | if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { |
| 4882 | sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); |
| 4883 | if (sk->sk_socket && |
| 4884 | test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) |
| 4885 | tcp_new_space(sk); |
| 4886 | } |
| 4887 | } |
| 4888 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 4889 | static inline void tcp_data_snd_check(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4890 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 4891 | tcp_push_pending_frames(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4892 | tcp_check_space(sk); |
| 4893 | } |
| 4894 | |
| 4895 | /* |
| 4896 | * Check if sending an ack is needed. |
| 4897 | */ |
| 4898 | static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) |
| 4899 | { |
| 4900 | struct tcp_sock *tp = tcp_sk(sk); |
| 4901 | |
| 4902 | /* More than one full frame received... */ |
Joe Perches | 9d4fb27 | 2009-11-23 10:41:23 -0800 | [diff] [blame] | 4903 | if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss && |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4904 | /* ... and right edge of window advances far enough. |
| 4905 | * (tcp_recvmsg() will send ACK otherwise). Or... |
| 4906 | */ |
Joe Perches | 9d4fb27 | 2009-11-23 10:41:23 -0800 | [diff] [blame] | 4907 | __tcp_select_window(sk) >= tp->rcv_wnd) || |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4908 | /* We ACK each frame or... */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4909 | tcp_in_quickack_mode(sk) || |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4910 | /* We have out of order data. */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4911 | (ofo_possible && skb_peek(&tp->out_of_order_queue))) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4912 | /* Then ack it now */ |
| 4913 | tcp_send_ack(sk); |
| 4914 | } else { |
| 4915 | /* Else, send delayed ack. */ |
| 4916 | tcp_send_delayed_ack(sk); |
| 4917 | } |
| 4918 | } |
| 4919 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 4920 | static inline void tcp_ack_snd_check(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4921 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4922 | if (!inet_csk_ack_scheduled(sk)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4923 | /* We sent a data segment already. */ |
| 4924 | return; |
| 4925 | } |
| 4926 | __tcp_ack_snd_check(sk, 1); |
| 4927 | } |
| 4928 | |
| 4929 | /* |
| 4930 | * This routine is only called when we have urgent data |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 4931 | * signaled. Its the 'slow' part of tcp_urg. It could be |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4932 | * moved inline now as tcp_urg is only called from one |
| 4933 | * place. We handle URGent data wrong. We have to - as |
| 4934 | * BSD still doesn't use the correction from RFC961. |
| 4935 | * For 1003.1g we should support a new option TCP_STDURG to permit |
| 4936 | * either form (or just set the sysctl tcp_stdurg). |
| 4937 | */ |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4938 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4939 | static void tcp_check_urg(struct sock *sk, struct tcphdr *th) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4940 | { |
| 4941 | struct tcp_sock *tp = tcp_sk(sk); |
| 4942 | u32 ptr = ntohs(th->urg_ptr); |
| 4943 | |
| 4944 | if (ptr && !sysctl_tcp_stdurg) |
| 4945 | ptr--; |
| 4946 | ptr += ntohl(th->seq); |
| 4947 | |
| 4948 | /* Ignore urgent data that we've already seen and read. */ |
| 4949 | if (after(tp->copied_seq, ptr)) |
| 4950 | return; |
| 4951 | |
| 4952 | /* Do not replay urg ptr. |
| 4953 | * |
| 4954 | * NOTE: interesting situation not covered by specs. |
| 4955 | * Misbehaving sender may send urg ptr, pointing to segment, |
| 4956 | * which we already have in ofo queue. We are not able to fetch |
| 4957 | * such data and will stay in TCP_URG_NOTYET until will be eaten |
| 4958 | * by recvmsg(). Seems, we are not obliged to handle such wicked |
| 4959 | * situations. But it is worth to think about possibility of some |
| 4960 | * DoSes using some hypothetical application level deadlock. |
| 4961 | */ |
| 4962 | if (before(ptr, tp->rcv_nxt)) |
| 4963 | return; |
| 4964 | |
| 4965 | /* Do we already have a newer (or duplicate) urgent pointer? */ |
| 4966 | if (tp->urg_data && !after(ptr, tp->urg_seq)) |
| 4967 | return; |
| 4968 | |
| 4969 | /* Tell the world about our new urgent pointer. */ |
| 4970 | sk_send_sigurg(sk); |
| 4971 | |
| 4972 | /* We may be adding urgent data when the last byte read was |
| 4973 | * urgent. To do this requires some care. We cannot just ignore |
| 4974 | * tp->copied_seq since we would read the last urgent byte again |
| 4975 | * as data, nor can we alter copied_seq until this data arrives |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 4976 | * or we break the semantics of SIOCATMARK (and thus sockatmark()) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4977 | * |
| 4978 | * NOTE. Double Dutch. Rendering to plain English: author of comment |
| 4979 | * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); |
| 4980 | * and expect that both A and B disappear from stream. This is _wrong_. |
| 4981 | * Though this happens in BSD with high probability, this is occasional. |
| 4982 | * Any application relying on this is buggy. Note also, that fix "works" |
| 4983 | * only in this artificial test. Insert some normal data between A and B and we will |
| 4984 | * decline of BSD again. Verdict: it is better to remove to trap |
| 4985 | * buggy users. |
| 4986 | */ |
| 4987 | if (tp->urg_seq == tp->copied_seq && tp->urg_data && |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4988 | !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4989 | struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); |
| 4990 | tp->copied_seq++; |
| 4991 | if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { |
David S. Miller | 8728b83 | 2005-08-09 19:25:21 -0700 | [diff] [blame] | 4992 | __skb_unlink(skb, &sk->sk_receive_queue); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4993 | __kfree_skb(skb); |
| 4994 | } |
| 4995 | } |
| 4996 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4997 | tp->urg_data = TCP_URG_NOTYET; |
| 4998 | tp->urg_seq = ptr; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4999 | |
| 5000 | /* Disable header prediction. */ |
| 5001 | tp->pred_flags = 0; |
| 5002 | } |
| 5003 | |
| 5004 | /* This is the 'fast' part of urgent handling. */ |
| 5005 | static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th) |
| 5006 | { |
| 5007 | struct tcp_sock *tp = tcp_sk(sk); |
| 5008 | |
| 5009 | /* Check if we get a new urgent pointer - normally not. */ |
| 5010 | if (th->urg) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 5011 | tcp_check_urg(sk, th); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5012 | |
| 5013 | /* Do we wait for any urgent data? - normally not... */ |
| 5014 | if (tp->urg_data == TCP_URG_NOTYET) { |
| 5015 | u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - |
| 5016 | th->syn; |
| 5017 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5018 | /* Is the urgent pointer pointing into this packet? */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5019 | if (ptr < skb->len) { |
| 5020 | u8 tmp; |
| 5021 | if (skb_copy_bits(skb, ptr, &tmp, 1)) |
| 5022 | BUG(); |
| 5023 | tp->urg_data = TCP_URG_VALID | tmp; |
| 5024 | if (!sock_flag(sk, SOCK_DEAD)) |
| 5025 | sk->sk_data_ready(sk, 0); |
| 5026 | } |
| 5027 | } |
| 5028 | } |
| 5029 | |
| 5030 | static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) |
| 5031 | { |
| 5032 | struct tcp_sock *tp = tcp_sk(sk); |
| 5033 | int chunk = skb->len - hlen; |
| 5034 | int err; |
| 5035 | |
| 5036 | local_bh_enable(); |
Herbert Xu | 6047637 | 2007-04-09 11:59:39 -0700 | [diff] [blame] | 5037 | if (skb_csum_unnecessary(skb)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5038 | err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); |
| 5039 | else |
| 5040 | err = skb_copy_and_csum_datagram_iovec(skb, hlen, |
| 5041 | tp->ucopy.iov); |
| 5042 | |
| 5043 | if (!err) { |
| 5044 | tp->ucopy.len -= chunk; |
| 5045 | tp->copied_seq += chunk; |
| 5046 | tcp_rcv_space_adjust(sk); |
| 5047 | } |
| 5048 | |
| 5049 | local_bh_disable(); |
| 5050 | return err; |
| 5051 | } |
| 5052 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 5053 | static __sum16 __tcp_checksum_complete_user(struct sock *sk, |
| 5054 | struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5055 | { |
Al Viro | b51655b | 2006-11-14 21:40:42 -0800 | [diff] [blame] | 5056 | __sum16 result; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5057 | |
| 5058 | if (sock_owned_by_user(sk)) { |
| 5059 | local_bh_enable(); |
| 5060 | result = __tcp_checksum_complete(skb); |
| 5061 | local_bh_disable(); |
| 5062 | } else { |
| 5063 | result = __tcp_checksum_complete(skb); |
| 5064 | } |
| 5065 | return result; |
| 5066 | } |
| 5067 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 5068 | static inline int tcp_checksum_complete_user(struct sock *sk, |
| 5069 | struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5070 | { |
Herbert Xu | 6047637 | 2007-04-09 11:59:39 -0700 | [diff] [blame] | 5071 | return !skb_csum_unnecessary(skb) && |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 5072 | __tcp_checksum_complete_user(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5073 | } |
| 5074 | |
Chris Leech | 1a2449a | 2006-05-23 18:05:53 -0700 | [diff] [blame] | 5075 | #ifdef CONFIG_NET_DMA |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 5076 | static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, |
| 5077 | int hlen) |
Chris Leech | 1a2449a | 2006-05-23 18:05:53 -0700 | [diff] [blame] | 5078 | { |
| 5079 | struct tcp_sock *tp = tcp_sk(sk); |
| 5080 | int chunk = skb->len - hlen; |
| 5081 | int dma_cookie; |
| 5082 | int copied_early = 0; |
| 5083 | |
| 5084 | if (tp->ucopy.wakeup) |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5085 | return 0; |
Chris Leech | 1a2449a | 2006-05-23 18:05:53 -0700 | [diff] [blame] | 5086 | |
| 5087 | if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) |
Dan Williams | f67b459 | 2009-01-06 11:38:15 -0700 | [diff] [blame] | 5088 | tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY); |
Chris Leech | 1a2449a | 2006-05-23 18:05:53 -0700 | [diff] [blame] | 5089 | |
Herbert Xu | 6047637 | 2007-04-09 11:59:39 -0700 | [diff] [blame] | 5090 | if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) { |
Chris Leech | 1a2449a | 2006-05-23 18:05:53 -0700 | [diff] [blame] | 5091 | |
| 5092 | dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 5093 | skb, hlen, |
| 5094 | tp->ucopy.iov, chunk, |
| 5095 | tp->ucopy.pinned_list); |
Chris Leech | 1a2449a | 2006-05-23 18:05:53 -0700 | [diff] [blame] | 5096 | |
| 5097 | if (dma_cookie < 0) |
| 5098 | goto out; |
| 5099 | |
| 5100 | tp->ucopy.dma_cookie = dma_cookie; |
| 5101 | copied_early = 1; |
| 5102 | |
| 5103 | tp->ucopy.len -= chunk; |
| 5104 | tp->copied_seq += chunk; |
| 5105 | tcp_rcv_space_adjust(sk); |
| 5106 | |
| 5107 | if ((tp->ucopy.len == 0) || |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 5108 | (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) || |
Chris Leech | 1a2449a | 2006-05-23 18:05:53 -0700 | [diff] [blame] | 5109 | (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) { |
| 5110 | tp->ucopy.wakeup = 1; |
| 5111 | sk->sk_data_ready(sk, 0); |
| 5112 | } |
| 5113 | } else if (chunk > 0) { |
| 5114 | tp->ucopy.wakeup = 1; |
| 5115 | sk->sk_data_ready(sk, 0); |
| 5116 | } |
| 5117 | out: |
| 5118 | return copied_early; |
| 5119 | } |
| 5120 | #endif /* CONFIG_NET_DMA */ |
| 5121 | |
Ilpo Järvinen | cbe2d12 | 2008-08-23 05:10:12 -0700 | [diff] [blame] | 5122 | /* Does PAWS and seqno based validation of an incoming segment, flags will |
| 5123 | * play significant role here. |
| 5124 | */ |
| 5125 | static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb, |
| 5126 | struct tcphdr *th, int syn_inerr) |
| 5127 | { |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 5128 | u8 *hash_location; |
Ilpo Järvinen | cbe2d12 | 2008-08-23 05:10:12 -0700 | [diff] [blame] | 5129 | struct tcp_sock *tp = tcp_sk(sk); |
| 5130 | |
| 5131 | /* RFC1323: H1. Apply PAWS check first. */ |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 5132 | if (tcp_fast_parse_options(skb, th, tp, &hash_location) && |
| 5133 | tp->rx_opt.saw_tstamp && |
Ilpo Järvinen | cbe2d12 | 2008-08-23 05:10:12 -0700 | [diff] [blame] | 5134 | tcp_paws_discard(sk, skb)) { |
| 5135 | if (!th->rst) { |
| 5136 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); |
| 5137 | tcp_send_dupack(sk, skb); |
| 5138 | goto discard; |
| 5139 | } |
| 5140 | /* Reset is accepted even if it did not pass PAWS. */ |
| 5141 | } |
| 5142 | |
| 5143 | /* Step 1: check sequence number */ |
| 5144 | if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { |
| 5145 | /* RFC793, page 37: "In all states except SYN-SENT, all reset |
| 5146 | * (RST) segments are validated by checking their SEQ-fields." |
| 5147 | * And page 69: "If an incoming segment is not acceptable, |
| 5148 | * an acknowledgment should be sent in reply (unless the RST |
| 5149 | * bit is set, if so drop the segment and return)". |
| 5150 | */ |
| 5151 | if (!th->rst) |
| 5152 | tcp_send_dupack(sk, skb); |
| 5153 | goto discard; |
| 5154 | } |
| 5155 | |
| 5156 | /* Step 2: check RST bit */ |
| 5157 | if (th->rst) { |
| 5158 | tcp_reset(sk); |
| 5159 | goto discard; |
| 5160 | } |
| 5161 | |
| 5162 | /* ts_recent update must be made after we are sure that the packet |
| 5163 | * is in window. |
| 5164 | */ |
| 5165 | tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); |
| 5166 | |
| 5167 | /* step 3: check security and precedence [ignored] */ |
| 5168 | |
| 5169 | /* step 4: Check for a SYN in window. */ |
| 5170 | if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { |
| 5171 | if (syn_inerr) |
| 5172 | TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); |
| 5173 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN); |
| 5174 | tcp_reset(sk); |
| 5175 | return -1; |
| 5176 | } |
| 5177 | |
| 5178 | return 1; |
| 5179 | |
| 5180 | discard: |
| 5181 | __kfree_skb(skb); |
| 5182 | return 0; |
| 5183 | } |
| 5184 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5185 | /* |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5186 | * TCP receive function for the ESTABLISHED state. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5187 | * |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5188 | * It is split into a fast path and a slow path. The fast path is |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5189 | * disabled when: |
| 5190 | * - A zero window was announced from us - zero window probing |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5191 | * is only handled properly in the slow path. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5192 | * - Out of order segments arrived. |
| 5193 | * - Urgent data is expected. |
| 5194 | * - There is no buffer space left |
| 5195 | * - Unexpected TCP flags/window values/header lengths are received |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5196 | * (detected by checking the TCP header against pred_flags) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5197 | * - Data is sent in both directions. Fast path only supports pure senders |
| 5198 | * or pure receivers (this means either the sequence number or the ack |
| 5199 | * value must stay constant) |
| 5200 | * - Unexpected TCP option. |
| 5201 | * |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5202 | * When these conditions are not satisfied it drops into a standard |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5203 | * receive procedure patterned after RFC793 to handle all cases. |
| 5204 | * The first three cases are guaranteed by proper pred_flags setting, |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5205 | * the rest is checked inline. Fast processing is turned on in |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5206 | * tcp_data_queue when everything is OK. |
| 5207 | */ |
| 5208 | int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, |
| 5209 | struct tcphdr *th, unsigned len) |
| 5210 | { |
| 5211 | struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | cbe2d12 | 2008-08-23 05:10:12 -0700 | [diff] [blame] | 5212 | int res; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5213 | |
| 5214 | /* |
| 5215 | * Header prediction. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5216 | * The code loosely follows the one in the famous |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5217 | * "30 instruction TCP receive" Van Jacobson mail. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5218 | * |
| 5219 | * Van's trick is to deposit buffers into socket queue |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5220 | * on a device interrupt, to call tcp_recv function |
| 5221 | * on the receive process context and checksum and copy |
| 5222 | * the buffer to user space. smart... |
| 5223 | * |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5224 | * Our current scheme is not silly either but we take the |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5225 | * extra cost of the net_bh soft interrupt processing... |
| 5226 | * We do checksum and copy also but from device to kernel. |
| 5227 | */ |
| 5228 | |
| 5229 | tp->rx_opt.saw_tstamp = 0; |
| 5230 | |
| 5231 | /* pred_flags is 0xS?10 << 16 + snd_wnd |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 5232 | * if header_prediction is to be made |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5233 | * 'S' will always be tp->tcp_header_len >> 2 |
| 5234 | * '?' will be 0 for the fast path, otherwise pred_flags is 0 to |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5235 | * turn it off (when there are holes in the receive |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5236 | * space for instance) |
| 5237 | * PSH flag is ignored. |
| 5238 | */ |
| 5239 | |
| 5240 | if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && |
John Dykstra | 96e0bf4 | 2009-03-22 21:49:57 -0700 | [diff] [blame] | 5241 | TCP_SKB_CB(skb)->seq == tp->rcv_nxt && |
| 5242 | !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5243 | int tcp_header_len = tp->tcp_header_len; |
| 5244 | |
| 5245 | /* Timestamp header prediction: tcp_header_len |
| 5246 | * is automatically equal to th->doff*4 due to pred_flags |
| 5247 | * match. |
| 5248 | */ |
| 5249 | |
| 5250 | /* Check timestamp */ |
| 5251 | if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5252 | /* No? Slow path! */ |
Ilpo Järvinen | a4356b2 | 2008-08-23 05:12:29 -0700 | [diff] [blame] | 5253 | if (!tcp_parse_aligned_timestamp(tp, th)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5254 | goto slow_path; |
| 5255 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5256 | /* If PAWS failed, check it more carefully in slow path */ |
| 5257 | if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) |
| 5258 | goto slow_path; |
| 5259 | |
| 5260 | /* DO NOT update ts_recent here, if checksum fails |
| 5261 | * and timestamp was corrupted part, it will result |
| 5262 | * in a hung connection since we will drop all |
| 5263 | * future packets due to the PAWS test. |
| 5264 | */ |
| 5265 | } |
| 5266 | |
| 5267 | if (len <= tcp_header_len) { |
| 5268 | /* Bulk data transfer: sender */ |
| 5269 | if (len == tcp_header_len) { |
| 5270 | /* Predicted packet is in window by definition. |
| 5271 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. |
| 5272 | * Hence, check seq<=rcv_wup reduces to: |
| 5273 | */ |
| 5274 | if (tcp_header_len == |
| 5275 | (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && |
| 5276 | tp->rcv_nxt == tp->rcv_wup) |
| 5277 | tcp_store_ts_recent(tp); |
| 5278 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5279 | /* We know that such packets are checksummed |
| 5280 | * on entry. |
| 5281 | */ |
| 5282 | tcp_ack(sk, skb, 0); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5283 | __kfree_skb(skb); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 5284 | tcp_data_snd_check(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5285 | return 0; |
| 5286 | } else { /* Header too small */ |
Pavel Emelyanov | 63231bd | 2008-07-16 20:22:25 -0700 | [diff] [blame] | 5287 | TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5288 | goto discard; |
| 5289 | } |
| 5290 | } else { |
| 5291 | int eaten = 0; |
Chris Leech | 1a2449a | 2006-05-23 18:05:53 -0700 | [diff] [blame] | 5292 | int copied_early = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5293 | |
Chris Leech | 1a2449a | 2006-05-23 18:05:53 -0700 | [diff] [blame] | 5294 | if (tp->copied_seq == tp->rcv_nxt && |
| 5295 | len - tcp_header_len <= tp->ucopy.len) { |
| 5296 | #ifdef CONFIG_NET_DMA |
| 5297 | if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) { |
| 5298 | copied_early = 1; |
| 5299 | eaten = 1; |
| 5300 | } |
| 5301 | #endif |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 5302 | if (tp->ucopy.task == current && |
| 5303 | sock_owned_by_user(sk) && !copied_early) { |
Chris Leech | 1a2449a | 2006-05-23 18:05:53 -0700 | [diff] [blame] | 5304 | __set_current_state(TASK_RUNNING); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5305 | |
Chris Leech | 1a2449a | 2006-05-23 18:05:53 -0700 | [diff] [blame] | 5306 | if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) |
| 5307 | eaten = 1; |
| 5308 | } |
| 5309 | if (eaten) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5310 | /* Predicted packet is in window by definition. |
| 5311 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. |
| 5312 | * Hence, check seq<=rcv_wup reduces to: |
| 5313 | */ |
| 5314 | if (tcp_header_len == |
| 5315 | (sizeof(struct tcphdr) + |
| 5316 | TCPOLEN_TSTAMP_ALIGNED) && |
| 5317 | tp->rcv_nxt == tp->rcv_wup) |
| 5318 | tcp_store_ts_recent(tp); |
| 5319 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 5320 | tcp_rcv_rtt_measure_ts(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5321 | |
| 5322 | __skb_pull(skb, tcp_header_len); |
| 5323 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 5324 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5325 | } |
Chris Leech | 1a2449a | 2006-05-23 18:05:53 -0700 | [diff] [blame] | 5326 | if (copied_early) |
| 5327 | tcp_cleanup_rbuf(sk, skb->len); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5328 | } |
| 5329 | if (!eaten) { |
| 5330 | if (tcp_checksum_complete_user(sk, skb)) |
| 5331 | goto csum_error; |
| 5332 | |
| 5333 | /* Predicted packet is in window by definition. |
| 5334 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. |
| 5335 | * Hence, check seq<=rcv_wup reduces to: |
| 5336 | */ |
| 5337 | if (tcp_header_len == |
| 5338 | (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && |
| 5339 | tp->rcv_nxt == tp->rcv_wup) |
| 5340 | tcp_store_ts_recent(tp); |
| 5341 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 5342 | tcp_rcv_rtt_measure_ts(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5343 | |
| 5344 | if ((int)skb->truesize > sk->sk_forward_alloc) |
| 5345 | goto step5; |
| 5346 | |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 5347 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5348 | |
| 5349 | /* Bulk data transfer: receiver */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 5350 | __skb_pull(skb, tcp_header_len); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5351 | __skb_queue_tail(&sk->sk_receive_queue, skb); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 5352 | skb_set_owner_r(skb, sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5353 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; |
| 5354 | } |
| 5355 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 5356 | tcp_event_data_recv(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5357 | |
| 5358 | if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { |
| 5359 | /* Well, only one small jumplet in fast path... */ |
| 5360 | tcp_ack(sk, skb, FLAG_DATA); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 5361 | tcp_data_snd_check(sk); |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 5362 | if (!inet_csk_ack_scheduled(sk)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5363 | goto no_ack; |
| 5364 | } |
| 5365 | |
Ali Saidi | 53240c2 | 2008-10-07 15:31:19 -0700 | [diff] [blame] | 5366 | if (!copied_early || tp->rcv_nxt != tp->rcv_wup) |
| 5367 | __tcp_ack_snd_check(sk, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5368 | no_ack: |
Chris Leech | 1a2449a | 2006-05-23 18:05:53 -0700 | [diff] [blame] | 5369 | #ifdef CONFIG_NET_DMA |
| 5370 | if (copied_early) |
| 5371 | __skb_queue_tail(&sk->sk_async_wait_queue, skb); |
| 5372 | else |
| 5373 | #endif |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5374 | if (eaten) |
| 5375 | __kfree_skb(skb); |
| 5376 | else |
| 5377 | sk->sk_data_ready(sk, 0); |
| 5378 | return 0; |
| 5379 | } |
| 5380 | } |
| 5381 | |
| 5382 | slow_path: |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 5383 | if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5384 | goto csum_error; |
| 5385 | |
| 5386 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5387 | * Standard slow path. |
| 5388 | */ |
| 5389 | |
Ilpo Järvinen | cbe2d12 | 2008-08-23 05:10:12 -0700 | [diff] [blame] | 5390 | res = tcp_validate_incoming(sk, skb, th, 1); |
| 5391 | if (res <= 0) |
| 5392 | return -res; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5393 | |
| 5394 | step5: |
John Dykstra | 96e0bf4 | 2009-03-22 21:49:57 -0700 | [diff] [blame] | 5395 | if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0) |
| 5396 | goto discard; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5397 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 5398 | tcp_rcv_rtt_measure_ts(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5399 | |
| 5400 | /* Process urgent data. */ |
| 5401 | tcp_urg(sk, skb, th); |
| 5402 | |
| 5403 | /* step 7: process the segment text */ |
| 5404 | tcp_data_queue(sk, skb); |
| 5405 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 5406 | tcp_data_snd_check(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5407 | tcp_ack_snd_check(sk); |
| 5408 | return 0; |
| 5409 | |
| 5410 | csum_error: |
Pavel Emelyanov | 63231bd | 2008-07-16 20:22:25 -0700 | [diff] [blame] | 5411 | TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5412 | |
| 5413 | discard: |
| 5414 | __kfree_skb(skb); |
| 5415 | return 0; |
| 5416 | } |
| 5417 | |
| 5418 | static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, |
| 5419 | struct tcphdr *th, unsigned len) |
| 5420 | { |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 5421 | u8 *hash_location; |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 5422 | struct inet_connection_sock *icsk = inet_csk(sk); |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 5423 | struct tcp_sock *tp = tcp_sk(sk); |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 5424 | struct tcp_cookie_values *cvp = tp->cookie_values; |
| 5425 | int saved_clamp = tp->rx_opt.mss_clamp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5426 | |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 5427 | tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5428 | |
| 5429 | if (th->ack) { |
| 5430 | /* rfc793: |
| 5431 | * "If the state is SYN-SENT then |
| 5432 | * first check the ACK bit |
| 5433 | * If the ACK bit is set |
| 5434 | * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send |
| 5435 | * a reset (unless the RST bit is set, if so drop |
| 5436 | * the segment and return)" |
| 5437 | * |
| 5438 | * We do not send data with SYN, so that RFC-correct |
| 5439 | * test reduces to: |
| 5440 | */ |
| 5441 | if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt) |
| 5442 | goto reset_and_undo; |
| 5443 | |
| 5444 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && |
| 5445 | !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, |
| 5446 | tcp_time_stamp)) { |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 5447 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5448 | goto reset_and_undo; |
| 5449 | } |
| 5450 | |
| 5451 | /* Now ACK is acceptable. |
| 5452 | * |
| 5453 | * "If the RST bit is set |
| 5454 | * If the ACK was acceptable then signal the user "error: |
| 5455 | * connection reset", drop the segment, enter CLOSED state, |
| 5456 | * delete TCB, and return." |
| 5457 | */ |
| 5458 | |
| 5459 | if (th->rst) { |
| 5460 | tcp_reset(sk); |
| 5461 | goto discard; |
| 5462 | } |
| 5463 | |
| 5464 | /* rfc793: |
| 5465 | * "fifth, if neither of the SYN or RST bits is set then |
| 5466 | * drop the segment and return." |
| 5467 | * |
| 5468 | * See note below! |
| 5469 | * --ANK(990513) |
| 5470 | */ |
| 5471 | if (!th->syn) |
| 5472 | goto discard_and_undo; |
| 5473 | |
| 5474 | /* rfc793: |
| 5475 | * "If the SYN bit is on ... |
| 5476 | * are acceptable then ... |
| 5477 | * (our SYN has been ACKed), change the connection |
| 5478 | * state to ESTABLISHED..." |
| 5479 | */ |
| 5480 | |
| 5481 | TCP_ECN_rcv_synack(tp, th); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5482 | |
| 5483 | tp->snd_wl1 = TCP_SKB_CB(skb)->seq; |
| 5484 | tcp_ack(sk, skb, FLAG_SLOWPATH); |
| 5485 | |
| 5486 | /* Ok.. it's good. Set up sequence numbers and |
| 5487 | * move to established. |
| 5488 | */ |
| 5489 | tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; |
| 5490 | tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; |
| 5491 | |
| 5492 | /* RFC1323: The window in SYN & SYN/ACK segments is |
| 5493 | * never scaled. |
| 5494 | */ |
| 5495 | tp->snd_wnd = ntohs(th->window); |
Hantzis Fotis | ee7537b | 2009-03-02 22:42:02 -0800 | [diff] [blame] | 5496 | tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5497 | |
| 5498 | if (!tp->rx_opt.wscale_ok) { |
| 5499 | tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; |
| 5500 | tp->window_clamp = min(tp->window_clamp, 65535U); |
| 5501 | } |
| 5502 | |
| 5503 | if (tp->rx_opt.saw_tstamp) { |
| 5504 | tp->rx_opt.tstamp_ok = 1; |
| 5505 | tp->tcp_header_len = |
| 5506 | sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; |
| 5507 | tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; |
| 5508 | tcp_store_ts_recent(tp); |
| 5509 | } else { |
| 5510 | tp->tcp_header_len = sizeof(struct tcphdr); |
| 5511 | } |
| 5512 | |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 5513 | if (tcp_is_sack(tp) && sysctl_tcp_fack) |
| 5514 | tcp_enable_fack(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5515 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 5516 | tcp_mtup_init(sk); |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 5517 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5518 | tcp_initialize_rcv_mss(sk); |
| 5519 | |
| 5520 | /* Remember, tcp_poll() does not lock socket! |
| 5521 | * Change state from SYN-SENT only after copied_seq |
| 5522 | * is initialized. */ |
| 5523 | tp->copied_seq = tp->rcv_nxt; |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 5524 | |
| 5525 | if (cvp != NULL && |
| 5526 | cvp->cookie_pair_size > 0 && |
| 5527 | tp->rx_opt.cookie_plus > 0) { |
| 5528 | int cookie_size = tp->rx_opt.cookie_plus |
| 5529 | - TCPOLEN_COOKIE_BASE; |
| 5530 | int cookie_pair_size = cookie_size |
| 5531 | + cvp->cookie_desired; |
| 5532 | |
| 5533 | /* A cookie extension option was sent and returned. |
| 5534 | * Note that each incoming SYNACK replaces the |
| 5535 | * Responder cookie. The initial exchange is most |
| 5536 | * fragile, as protection against spoofing relies |
| 5537 | * entirely upon the sequence and timestamp (above). |
| 5538 | * This replacement strategy allows the correct pair to |
| 5539 | * pass through, while any others will be filtered via |
| 5540 | * Responder verification later. |
| 5541 | */ |
| 5542 | if (sizeof(cvp->cookie_pair) >= cookie_pair_size) { |
| 5543 | memcpy(&cvp->cookie_pair[cvp->cookie_desired], |
| 5544 | hash_location, cookie_size); |
| 5545 | cvp->cookie_pair_size = cookie_pair_size; |
| 5546 | } |
| 5547 | } |
| 5548 | |
Ralf Baechle | e16aa20 | 2006-12-07 00:11:33 -0800 | [diff] [blame] | 5549 | smp_mb(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5550 | tcp_set_state(sk, TCP_ESTABLISHED); |
| 5551 | |
Venkat Yekkirala | 6b87769 | 2006-11-08 17:04:09 -0600 | [diff] [blame] | 5552 | security_inet_conn_established(sk, skb); |
| 5553 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5554 | /* Make sure socket is routed, for correct metrics. */ |
Arnaldo Carvalho de Melo | 8292a17 | 2005-12-13 23:15:52 -0800 | [diff] [blame] | 5555 | icsk->icsk_af_ops->rebuild_header(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5556 | |
| 5557 | tcp_init_metrics(sk); |
| 5558 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 5559 | tcp_init_congestion_control(sk); |
Stephen Hemminger | 317a76f | 2005-06-23 12:19:55 -0700 | [diff] [blame] | 5560 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5561 | /* Prevent spurious tcp_cwnd_restart() on first data |
| 5562 | * packet. |
| 5563 | */ |
| 5564 | tp->lsndtime = tcp_time_stamp; |
| 5565 | |
| 5566 | tcp_init_buffer_space(sk); |
| 5567 | |
| 5568 | if (sock_flag(sk, SOCK_KEEPOPEN)) |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 5569 | inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5570 | |
| 5571 | if (!tp->rx_opt.snd_wscale) |
| 5572 | __tcp_fast_path_on(tp, tp->snd_wnd); |
| 5573 | else |
| 5574 | tp->pred_flags = 0; |
| 5575 | |
| 5576 | if (!sock_flag(sk, SOCK_DEAD)) { |
| 5577 | sk->sk_state_change(sk); |
Pavel Emelyanov | 8d8ad9d | 2007-11-26 20:10:50 +0800 | [diff] [blame] | 5578 | sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5579 | } |
| 5580 | |
Arnaldo Carvalho de Melo | 295f732 | 2005-08-09 20:11:56 -0700 | [diff] [blame] | 5581 | if (sk->sk_write_pending || |
| 5582 | icsk->icsk_accept_queue.rskq_defer_accept || |
| 5583 | icsk->icsk_ack.pingpong) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5584 | /* Save one ACK. Data will be ready after |
| 5585 | * several ticks, if write_pending is set. |
| 5586 | * |
| 5587 | * It may be deleted, but with this feature tcpdumps |
| 5588 | * look so _wonderfully_ clever, that I was not able |
| 5589 | * to stand against the temptation 8) --ANK |
| 5590 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 5591 | inet_csk_schedule_ack(sk); |
Arnaldo Carvalho de Melo | 295f732 | 2005-08-09 20:11:56 -0700 | [diff] [blame] | 5592 | icsk->icsk_ack.lrcvtime = tcp_time_stamp; |
| 5593 | icsk->icsk_ack.ato = TCP_ATO_MIN; |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 5594 | tcp_incr_quickack(sk); |
| 5595 | tcp_enter_quickack_mode(sk); |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 5596 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, |
| 5597 | TCP_DELACK_MAX, TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5598 | |
| 5599 | discard: |
| 5600 | __kfree_skb(skb); |
| 5601 | return 0; |
| 5602 | } else { |
| 5603 | tcp_send_ack(sk); |
| 5604 | } |
| 5605 | return -1; |
| 5606 | } |
| 5607 | |
| 5608 | /* No ACK in the segment */ |
| 5609 | |
| 5610 | if (th->rst) { |
| 5611 | /* rfc793: |
| 5612 | * "If the RST bit is set |
| 5613 | * |
| 5614 | * Otherwise (no ACK) drop the segment and return." |
| 5615 | */ |
| 5616 | |
| 5617 | goto discard_and_undo; |
| 5618 | } |
| 5619 | |
| 5620 | /* PAWS check. */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 5621 | if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && |
Ilpo Järvinen | c887e6d | 2009-03-14 14:23:03 +0000 | [diff] [blame] | 5622 | tcp_paws_reject(&tp->rx_opt, 0)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5623 | goto discard_and_undo; |
| 5624 | |
| 5625 | if (th->syn) { |
| 5626 | /* We see SYN without ACK. It is attempt of |
| 5627 | * simultaneous connect with crossed SYNs. |
| 5628 | * Particularly, it can be connect to self. |
| 5629 | */ |
| 5630 | tcp_set_state(sk, TCP_SYN_RECV); |
| 5631 | |
| 5632 | if (tp->rx_opt.saw_tstamp) { |
| 5633 | tp->rx_opt.tstamp_ok = 1; |
| 5634 | tcp_store_ts_recent(tp); |
| 5635 | tp->tcp_header_len = |
| 5636 | sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; |
| 5637 | } else { |
| 5638 | tp->tcp_header_len = sizeof(struct tcphdr); |
| 5639 | } |
| 5640 | |
| 5641 | tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; |
| 5642 | tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; |
| 5643 | |
| 5644 | /* RFC1323: The window in SYN & SYN/ACK segments is |
| 5645 | * never scaled. |
| 5646 | */ |
| 5647 | tp->snd_wnd = ntohs(th->window); |
| 5648 | tp->snd_wl1 = TCP_SKB_CB(skb)->seq; |
| 5649 | tp->max_window = tp->snd_wnd; |
| 5650 | |
| 5651 | TCP_ECN_rcv_syn(tp, th); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5652 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 5653 | tcp_mtup_init(sk); |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 5654 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5655 | tcp_initialize_rcv_mss(sk); |
| 5656 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5657 | tcp_send_synack(sk); |
| 5658 | #if 0 |
| 5659 | /* Note, we could accept data and URG from this segment. |
| 5660 | * There are no obstacles to make this. |
| 5661 | * |
| 5662 | * However, if we ignore data in ACKless segments sometimes, |
| 5663 | * we have no reasons to accept it sometimes. |
| 5664 | * Also, seems the code doing it in step6 of tcp_rcv_state_process |
| 5665 | * is not flawless. So, discard packet for sanity. |
| 5666 | * Uncomment this return to process the data. |
| 5667 | */ |
| 5668 | return -1; |
| 5669 | #else |
| 5670 | goto discard; |
| 5671 | #endif |
| 5672 | } |
| 5673 | /* "fifth, if neither of the SYN or RST bits is set then |
| 5674 | * drop the segment and return." |
| 5675 | */ |
| 5676 | |
| 5677 | discard_and_undo: |
| 5678 | tcp_clear_options(&tp->rx_opt); |
| 5679 | tp->rx_opt.mss_clamp = saved_clamp; |
| 5680 | goto discard; |
| 5681 | |
| 5682 | reset_and_undo: |
| 5683 | tcp_clear_options(&tp->rx_opt); |
| 5684 | tp->rx_opt.mss_clamp = saved_clamp; |
| 5685 | return 1; |
| 5686 | } |
| 5687 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5688 | /* |
| 5689 | * This function implements the receiving procedure of RFC 793 for |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5690 | * all states except ESTABLISHED and TIME_WAIT. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5691 | * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be |
| 5692 | * address independent. |
| 5693 | */ |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5694 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5695 | int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, |
| 5696 | struct tcphdr *th, unsigned len) |
| 5697 | { |
| 5698 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | 8292a17 | 2005-12-13 23:15:52 -0800 | [diff] [blame] | 5699 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5700 | int queued = 0; |
Ilpo Järvinen | cbe2d12 | 2008-08-23 05:10:12 -0700 | [diff] [blame] | 5701 | int res; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5702 | |
| 5703 | tp->rx_opt.saw_tstamp = 0; |
| 5704 | |
| 5705 | switch (sk->sk_state) { |
| 5706 | case TCP_CLOSE: |
| 5707 | goto discard; |
| 5708 | |
| 5709 | case TCP_LISTEN: |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 5710 | if (th->ack) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5711 | return 1; |
| 5712 | |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 5713 | if (th->rst) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5714 | goto discard; |
| 5715 | |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 5716 | if (th->syn) { |
Arnaldo Carvalho de Melo | 8292a17 | 2005-12-13 23:15:52 -0800 | [diff] [blame] | 5717 | if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5718 | return 1; |
| 5719 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5720 | /* Now we have several options: In theory there is |
| 5721 | * nothing else in the frame. KA9Q has an option to |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5722 | * send data with the syn, BSD accepts data with the |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5723 | * syn up to the [to be] advertised window and |
| 5724 | * Solaris 2.1 gives you a protocol error. For now |
| 5725 | * we just ignore it, that fits the spec precisely |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5726 | * and avoids incompatibilities. It would be nice in |
| 5727 | * future to drop through and process the data. |
| 5728 | * |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5729 | * Now that TTCP is starting to be used we ought to |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5730 | * queue this data. |
| 5731 | * But, this leaves one open to an easy denial of |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5732 | * service attack, and SYN cookies can't defend |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5733 | * against this problem. So, we drop the data |
Masayuki Nakagawa | fb7e239 | 2007-01-23 20:15:06 -0800 | [diff] [blame] | 5734 | * in the interest of security over speed unless |
| 5735 | * it's still in use. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5736 | */ |
Masayuki Nakagawa | fb7e239 | 2007-01-23 20:15:06 -0800 | [diff] [blame] | 5737 | kfree_skb(skb); |
| 5738 | return 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5739 | } |
| 5740 | goto discard; |
| 5741 | |
| 5742 | case TCP_SYN_SENT: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5743 | queued = tcp_rcv_synsent_state_process(sk, skb, th, len); |
| 5744 | if (queued >= 0) |
| 5745 | return queued; |
| 5746 | |
| 5747 | /* Do step6 onward by hand. */ |
| 5748 | tcp_urg(sk, skb, th); |
| 5749 | __kfree_skb(skb); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 5750 | tcp_data_snd_check(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5751 | return 0; |
| 5752 | } |
| 5753 | |
Ilpo Järvinen | cbe2d12 | 2008-08-23 05:10:12 -0700 | [diff] [blame] | 5754 | res = tcp_validate_incoming(sk, skb, th, 0); |
| 5755 | if (res <= 0) |
| 5756 | return -res; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5757 | |
| 5758 | /* step 5: check the ACK field */ |
| 5759 | if (th->ack) { |
John Dykstra | 96e0bf4 | 2009-03-22 21:49:57 -0700 | [diff] [blame] | 5760 | int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5761 | |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 5762 | switch (sk->sk_state) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5763 | case TCP_SYN_RECV: |
| 5764 | if (acceptable) { |
| 5765 | tp->copied_seq = tp->rcv_nxt; |
Ralf Baechle | e16aa20 | 2006-12-07 00:11:33 -0800 | [diff] [blame] | 5766 | smp_mb(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5767 | tcp_set_state(sk, TCP_ESTABLISHED); |
| 5768 | sk->sk_state_change(sk); |
| 5769 | |
| 5770 | /* Note, that this wakeup is only for marginal |
| 5771 | * crossed SYN case. Passively open sockets |
| 5772 | * are not waked up, because sk->sk_sleep == |
| 5773 | * NULL and sk->sk_socket == NULL. |
| 5774 | */ |
Pavel Emelyanov | 8d8ad9d | 2007-11-26 20:10:50 +0800 | [diff] [blame] | 5775 | if (sk->sk_socket) |
| 5776 | sk_wake_async(sk, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 5777 | SOCK_WAKE_IO, POLL_OUT); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5778 | |
| 5779 | tp->snd_una = TCP_SKB_CB(skb)->ack_seq; |
| 5780 | tp->snd_wnd = ntohs(th->window) << |
| 5781 | tp->rx_opt.snd_wscale; |
Hantzis Fotis | ee7537b | 2009-03-02 22:42:02 -0800 | [diff] [blame] | 5782 | tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5783 | |
| 5784 | /* tcp_ack considers this ACK as duplicate |
| 5785 | * and does not calculate rtt. |
Damian Lukowski | 5988564 | 2010-02-10 18:04:08 -0800 | [diff] [blame] | 5786 | * Force it here. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5787 | */ |
Damian Lukowski | 5988564 | 2010-02-10 18:04:08 -0800 | [diff] [blame] | 5788 | tcp_ack_update_rtt(sk, 0, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5789 | |
| 5790 | if (tp->rx_opt.tstamp_ok) |
| 5791 | tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; |
| 5792 | |
| 5793 | /* Make sure socket is routed, for |
| 5794 | * correct metrics. |
| 5795 | */ |
Arnaldo Carvalho de Melo | 8292a17 | 2005-12-13 23:15:52 -0800 | [diff] [blame] | 5796 | icsk->icsk_af_ops->rebuild_header(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5797 | |
| 5798 | tcp_init_metrics(sk); |
| 5799 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 5800 | tcp_init_congestion_control(sk); |
Stephen Hemminger | 317a76f | 2005-06-23 12:19:55 -0700 | [diff] [blame] | 5801 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5802 | /* Prevent spurious tcp_cwnd_restart() on |
| 5803 | * first data packet. |
| 5804 | */ |
| 5805 | tp->lsndtime = tcp_time_stamp; |
| 5806 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 5807 | tcp_mtup_init(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5808 | tcp_initialize_rcv_mss(sk); |
| 5809 | tcp_init_buffer_space(sk); |
| 5810 | tcp_fast_path_on(tp); |
| 5811 | } else { |
| 5812 | return 1; |
| 5813 | } |
| 5814 | break; |
| 5815 | |
| 5816 | case TCP_FIN_WAIT1: |
| 5817 | if (tp->snd_una == tp->write_seq) { |
| 5818 | tcp_set_state(sk, TCP_FIN_WAIT2); |
| 5819 | sk->sk_shutdown |= SEND_SHUTDOWN; |
| 5820 | dst_confirm(sk->sk_dst_cache); |
| 5821 | |
| 5822 | if (!sock_flag(sk, SOCK_DEAD)) |
| 5823 | /* Wake up lingering close() */ |
| 5824 | sk->sk_state_change(sk); |
| 5825 | else { |
| 5826 | int tmo; |
| 5827 | |
| 5828 | if (tp->linger2 < 0 || |
| 5829 | (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && |
| 5830 | after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { |
| 5831 | tcp_done(sk); |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 5832 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5833 | return 1; |
| 5834 | } |
| 5835 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 5836 | tmo = tcp_fin_time(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5837 | if (tmo > TCP_TIMEWAIT_LEN) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 5838 | inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5839 | } else if (th->fin || sock_owned_by_user(sk)) { |
| 5840 | /* Bad case. We could lose such FIN otherwise. |
| 5841 | * It is not a big problem, but it looks confusing |
| 5842 | * and not so rare event. We still can lose it now, |
| 5843 | * if it spins in bh_lock_sock(), but it is really |
| 5844 | * marginal case. |
| 5845 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 5846 | inet_csk_reset_keepalive_timer(sk, tmo); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5847 | } else { |
| 5848 | tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); |
| 5849 | goto discard; |
| 5850 | } |
| 5851 | } |
| 5852 | } |
| 5853 | break; |
| 5854 | |
| 5855 | case TCP_CLOSING: |
| 5856 | if (tp->snd_una == tp->write_seq) { |
| 5857 | tcp_time_wait(sk, TCP_TIME_WAIT, 0); |
| 5858 | goto discard; |
| 5859 | } |
| 5860 | break; |
| 5861 | |
| 5862 | case TCP_LAST_ACK: |
| 5863 | if (tp->snd_una == tp->write_seq) { |
| 5864 | tcp_update_metrics(sk); |
| 5865 | tcp_done(sk); |
| 5866 | goto discard; |
| 5867 | } |
| 5868 | break; |
| 5869 | } |
| 5870 | } else |
| 5871 | goto discard; |
| 5872 | |
| 5873 | /* step 6: check the URG bit */ |
| 5874 | tcp_urg(sk, skb, th); |
| 5875 | |
| 5876 | /* step 7: process the segment text */ |
| 5877 | switch (sk->sk_state) { |
| 5878 | case TCP_CLOSE_WAIT: |
| 5879 | case TCP_CLOSING: |
| 5880 | case TCP_LAST_ACK: |
| 5881 | if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) |
| 5882 | break; |
| 5883 | case TCP_FIN_WAIT1: |
| 5884 | case TCP_FIN_WAIT2: |
| 5885 | /* RFC 793 says to queue data in these states, |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5886 | * RFC 1122 says we MUST send a reset. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5887 | * BSD 4.4 also does reset. |
| 5888 | */ |
| 5889 | if (sk->sk_shutdown & RCV_SHUTDOWN) { |
| 5890 | if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && |
| 5891 | after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 5892 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5893 | tcp_reset(sk); |
| 5894 | return 1; |
| 5895 | } |
| 5896 | } |
| 5897 | /* Fall through */ |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5898 | case TCP_ESTABLISHED: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5899 | tcp_data_queue(sk, skb); |
| 5900 | queued = 1; |
| 5901 | break; |
| 5902 | } |
| 5903 | |
| 5904 | /* tcp_data could move socket to TIME-WAIT */ |
| 5905 | if (sk->sk_state != TCP_CLOSE) { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 5906 | tcp_data_snd_check(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5907 | tcp_ack_snd_check(sk); |
| 5908 | } |
| 5909 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5910 | if (!queued) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5911 | discard: |
| 5912 | __kfree_skb(skb); |
| 5913 | } |
| 5914 | return 0; |
| 5915 | } |
| 5916 | |
| 5917 | EXPORT_SYMBOL(sysctl_tcp_ecn); |
| 5918 | EXPORT_SYMBOL(sysctl_tcp_reordering); |
Glenn Griffin | c6aefaf | 2008-02-07 21:49:26 -0800 | [diff] [blame] | 5919 | EXPORT_SYMBOL(sysctl_tcp_adv_win_scale); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5920 | EXPORT_SYMBOL(tcp_parse_options); |
YOSHIFUJI Hideaki | 7d5d552 | 2008-04-17 12:29:53 +0900 | [diff] [blame] | 5921 | #ifdef CONFIG_TCP_MD5SIG |
| 5922 | EXPORT_SYMBOL(tcp_parse_md5sig_option); |
| 5923 | #endif |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5924 | EXPORT_SYMBOL(tcp_rcv_established); |
| 5925 | EXPORT_SYMBOL(tcp_rcv_state_process); |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 5926 | EXPORT_SYMBOL(tcp_initialize_rcv_mss); |